ProductUnhunt

Fake Vacation

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

How It Started

Travel is now two products pretending to be one. There is the actual trip — flights, queues, leave approvals, bad hotel pillows, expensive food, and the minor administrative burden of physically being somewhere. Then there is the version that survives: six photographs, three stories, one airport coffee, and enough evidence for everyone else to conclude that you have a life.

At some point the obvious optimisation presented itself: remove the expensive half.

The Pitch

Fake Vacation is a complete synthetic travel service for people who want the social outcome of a holiday without taking the holiday.

You pick a destination, dates, travel style, and how aggressively you normally post. You upload a few photos of yourself. Fake Vacation then creates the entire trip around you.

Not just twenty attractive AI images. That would be lazy.

The system researches the actual flight schedule, airport, departure time, arrival window, seasonal weather, sunrise and sunset, local events, crowd patterns, geography, and how long it would realistically take to move between places.

If your fictional Delhi-to-Tokyo flight lands at 07:10 local time, your first airport photo should look like 07:10 in Tokyo. If it rained in Shibuya that afternoon, you do not get a suspiciously perfect blue-sky crossing photo. If day three is colder than day two, you should probably not be wearing the same linen shirt.

The itinerary is generated day by day, and every artifact has to agree with the story.

The boring photos matter most. Hotel corridors. A blurry metro shot. A badly framed breakfast. A vending machine. A photo of nothing in particular because real vacations contain a statistically significant amount of photographic garbage.

And once the vacation is over, Fake Vacation gives you the part no image model remembers to generate:

The Vacation Debrief

A ten-minute briefing on the trip you did not take.

Where you stayed. What the neighbourhood was like. Which places you supposedly preferred. What the weather did. Which restaurant was overrated. What was surprisingly expensive. How the trains worked. What you would “definitely do differently next time.”

The goal is simple: when someone asks, “How was Lisbon?”, you should not answer, “Very… Portuguese.”

In the fictional pricing table, the arithmetic looked decisive:

Actual trip: ₹200,000. Synthetic trip: ₹2,000.

Same social proof. Considerably less jet lag.

The Delusion

For about eleven minutes this sounds unreasonably good.

Travel is expensive. Leave is scarce. Social media has already separated the experience from the evidence of the experience. The product merely completes the abstraction.

The unit economics are beautiful. Image generation is cheap. Itinerary generation is cheap. Research is automated. Storage is trivial. Gross margins look excellent. The customer acquires a week of lifestyle content for less than the price of one airport meal.

The product also appears to contain its own distribution engine. Every customer posts the output publicly. Every generated vacation is theoretically an advertisement.

Someone says “aspirational identity infrastructure.” Nobody stops the meeting.

The Reality Check

Fake Vacation has an impossible distribution problem.

If people know the service exists, the product stops working.

The moment Fake Vacation becomes recognisable, every airport photo becomes suspicious. Every improbable Maldives trip gets mentally reverse-engineered. Every customer starts wondering whether everyone else’s vacation is fake too.

The company’s success destroys the product’s value.

Unfortunately, the alternative is worse.

If nobody knows Fake Vacation exists, nobody buys Fake Vacation.

So the growth model is:

Stay obscure enough to preserve credibility.
Become visible enough to acquire customers.
Immediately damage the credibility required by step one.

The viral loop is not a growth loop. It is a self-destruct mechanism.

Then the smaller problems arrive.

Your mother asks how Lisbon was. A colleague was in Lisbon that week. Someone notices the museum you “visited” was closed on Tuesdays. You accidentally post a sunny afternoon from a day that had six hours of rain. Your generated hotel balcony appears in somebody else’s fake honeymoon.

The image quality can be solved. The continuity can be solved. The research can be solved.

The central contradiction cannot.

The product only works if it is undiscovered, and an undiscovered consumer product is generally known by another name: a failed business.

Why It Stays Unbuilt

Because the best-case scenario is that Fake Vacation becomes popular enough to make Fake Vacation useless.

Unwanted Bonus

Fake Sabbatical.

Six months of synthetic personal growth for people who cannot take six months off. Gradually generated hiking photos, reading lists, vague reflections about slowing down, one tasteful cabin, two carefully timed posts about “re-evaluating what matters,” and a final LinkedIn essay announcing that you have returned with clarity.

The premium tier includes a completely fabricated personality change.

I thought about building this and chose not to.

Churnstile

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

The Pitch

Every SaaS company builds two flows with wildly different budgets: the signup, lovingly A/B-tested for years, and the cancellation, written by an intern in 2019 and never opened again. Churnstile is the drop-in fix — a hosted cancellation flow with a tasteful exit survey, tiered save-offers, a pause plan, and an automated win-back sequence. Pricing is a platform fee plus a percentage of “saved revenue,” so we only make money when your customers stay.

Integration is one script tag. Benchmarks across the whole customer network tell you whether your churn is normal for your category. Nobody has this data.

The Delusion

This one is defensible, which is the problem. Churn is universal, the cancellation page is universally neglected, and “we take a cut of the revenue we save” is the kind of aligned-incentive pricing that makes an investor reach for a term sheet mid-sentence. The cross-network benchmark data is a real moat. There is an actual, honest version of this product: one that surfaces a pause plan to someone who just needed a quiet month, and lets everyone else leave with dignity.

We got four slides into the deck before anyone spotted the trapdoor.

The Reality Check

The trapdoor: a percentage of saved revenue means Churnstile earns the most when cancelling is hardest. The drift into dark patterns is not a risk to be managed — it is the pricing model, operating as designed. Every quarterly review becomes a meeting about making one more screen appear between a tired customer and the button they came for.

The legal shortcut is not as neat as the pitch deck claimed. The US Federal Trade Commission’s 2024 Click-to-Cancel rule was vacated before taking effect; in March 2026, the agency opened a new negative-option rulemaking. That uncertainty makes every extra screen a future compliance meeting with analytics attached.

And the recursion is unkind: Churnstile’s own customers will eventually want to cancel Churnstile, at which point our cancellation flow is either a hypocrisy or a live demo of everything they’re escaping.

Why It Stays Unbuilt

Because the good version gets absorbed as a checkbox feature by every billing platform within eighteen months, and the bad version ends with our exit survey being read aloud in a deposition. There is no durable middle. The moat was never the benchmark data; it was a willingness to be the last unpleasant thing a customer experiences on the way out, and that is not a moat anyone should want to defend.

Unwanted Bonus

By the end of the afternoon someone had sketched “retention insurance” — actuarially priced churn protection, underwritten with the benchmark data. It is a genuinely worse idea, and half the room wanted to pivot to it immediately. That was the moment we closed the whiteboard.

I thought about building this and chose not to.

Ambient Guilt

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

The Pitch

Calm technology, finally done right. No apps, no feeds, no notifications — just a hand-finished walnut frame, a museum-grade e-ink panel, and a single number, updated gently through the day: your unread email count. It hangs where a small painting would. Ambient, glanceable, honest. The website says “information as furniture.” The renders are, in fairness, gorgeous. $149, gift-boxed, with a linen cable.

The Delusion

Calm tech is having a moment, e-ink hardware margins are real, and the one-number purity gives it that Kickstarter-fatal combination of extremely easy to explain and extremely easy to film. It photographs like design press bait. It IS design press bait. “It’s like a weather station, but for your obligations” is the sort of line that closes a conversation. Direction unclear.

The Reality Check

We built an anxiety appliance and framed it in walnut. The number only goes up. It goes up during dinner. It goes up, visibly, in the background of the video call you’re taking from the sofa, where your guests, your partner, and your own peripheral vision can all watch your obligations compound in museum-grade grayscale. The one focus-group participant we still think about takes it off the wall, unplugs it, and puts it face-down in a drawer. This is an imagined test, but the drawer is still the most honest interface in the design.

Also it’s hardware, so that number ships with a supply chain, a firmware update pipeline, and a companion app for a product whose entire feature list is “a number.” The setup flow requires OAuth. Into your inbox. From a picture frame.

Why It Stays Unbuilt

Because the product works precisely as designed, and that is the problem. The brief said “make email ambient,” and email’s ambient form turns out to be dread. Some information became invisible for a reason; the drawer was the correct interface all along.

Unwanted Bonus

The roadmap listed Ambient Guilt Pro, which adds a second number. Slack. The projected attach rate was excellent, and we are confident it would have ended marriages.

I thought about building this and chose not to.

Doze Division

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

The Pitch

Your watch already scores your sleep. Doze Division makes the score mean something: weekly leagues of twenty strangers, promotion and relegation, seasonal playoffs, a global ladder. Streak bonuses for consistent bedtimes. Rivalries. A highlight reel, generated every morning, of a thing you cannot remember doing. Step-counting built empires on numbers this thin — and unlike steps, sleep requires no effort whatsoever. The most accessible sport ever invented. Everyone already trains eight hours a day.

The Delusion

Gamification turned walking, a thing humans did anyway, into a decade of shipped product. Sleep is a bigger market than walking — universal, measured, emotionally loaded — and it comes pre-instrumented, since the wearables are already on the wrists. Zero-effort onboarding, daily engagement by definition, and a leaderboard culture ready to bolt on. The pitch survives an astonishing amount of scrutiny before it collapses.

The Reality Check

It collapses on contact with the actual sport. Competitive pressure about sleep can destroy sleep. In a 2017 case report, researchers coined orthosomnia for an unhealthy preoccupation with improving sleep data from wearables. It is not a diagnosis or proof that every tracker harms sleep. It is, however, an excellent warning against adding playoffs. Nothing ruins a Tuesday night like needing eight competitive hours to stay in Division 2, checking the rank at 1 a.m., and watching your own anxiety relegate you in real time.

Cheating, meanwhile, requires no skill: the watch sleeps beautifully on a pillow. And our best athletes are definitionally unconscious during their achievements — they can’t stream, can’t spectate, can’t even watch themselves win.

Why It Stays Unbuilt

Because retention requires making users worse at the one thing the product measures. The harder someone tries to win Doze Division, the worse they sleep, and the worse they sleep, the more they open the app. That is a growth loop. It is also a medical complaint with a login screen.

Unwanted Bonus

The design file contained a spectator mode: live feeds of heart-rate graphs during the playoff finals, with commentary. Someone had mocked up a caster saying “he’s entering REM early — a bold strategy.” We closed the laptop gently, the way you’d put something to sleep.

I thought about building this and chose not to.

MacroLog

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

How It Started

I was using one of those AI-enabled calorie tracking apps. You photograph your dal, the AI tells you it is dal, and everyone feels very futuristic. On day seven the trial ended and the app asked for ₹700 a month — seven hundred Indian rupees.

₹700. A month. For inference.

I did what any emotionally healthy person does when confronted with a subscription: I opened a calculator. For my usage, I estimated that cheaper models on OpenRouter would cost roughly one US dollar for the entire month. And I eat more or less the same four things every day — embed them once in pgvector and most days the “AI” doesn’t even need to run. Why was I paying a convenience premium to call the same API?

This is the kind of question a normal person answers with “convenience” and moves on with their life. I answered it with a Laravel project.

The Pitch

MacroLog is calorie tracking for engineers — people who look at a ₹700 monthly app and see their own tiny API bill wearing a nice UI as a disguise.

You bring your own API key. You describe what you ate; a cheap model does the macro estimation; the bill goes to your OpenRouter account, where you can watch it fail to reach a dollar. Your meals get embedded in pgvector, so the second time you log “2 rotis and dal” — which, statistically speaking, is every time — it’s a vector lookup, not an API call. The system learns the profoundly repetitive truth about your diet and stops paying to rediscover it.

And surely I wasn’t alone? Surely there was a whole population of software developers out there, tracking their protein, silently resenting their subscriptions, waiting for a platform where they could do it at cost? People who read “bring your own key” and feel seen?

So I actually launched macrolog.online. It is live right now. It has one registered user. His last entry was six months ago. I can name him.

The name is not exclusive; other, unrelated nutrition products also call themselves MacroLog. This one is mine, which is the less reassuring distinction.

The Delusion

The estimate was genuinely persuasive — that’s the trap. For my repetitive diet, the pgvector caching trick works, and “BYOK so the platform can’t rug you on pricing” is a principle I still believe in.

The delusion was every word after “for myself.” I took “I don’t want to pay ₹700” and heard “there is a market of people who would rather configure an API key than pay ₹700.” Do you know what those two statements have in common? Nothing. The set of people who track calories and the set of people who know what OpenRouter is do overlap — I’ve met the overlap, it’s me — but “I am the target user” is market research the way a mirror is a focus group.

Then, when even that one user stopped logging in, the open source enthusiast in me woke up. Hell with third-party apps entirely — I put it on GitHub as laravel-macro-log, so that other people could self-host the thing I had already stopped using. This felt like a principled act at the time. It was actually a very elaborate way of closing a tab.

The Reality Check

Here is the entire go-to-market strategy, reproduced in full:

That’s it. That was the strategy. I built it, launched it, and waited for the developers-who-count-macros to arrive through what I can only assume was going to be word of mouth among people I had not told.

Nobody knows MacroLog exists. There is no distribution, no content, no community, no reason anyone searching “calorie tracker” would ever land on it. And the audience it’s for is precisely the audience least likely to pay for it — the pitch is literally “you could do this yourself for a dollar,” delivered to people who respond, “you’re right,” and then do it themselves for a dollar. I had built a product whose value proposition was an instruction manual for not needing the product.

Meanwhile the ₹700 app is doing fine, because its users want to photograph their dal, not provision credentials. The markup I was so offended by wasn’t inference cost. It was the price of never having to learn what an API key is. That, it turns out, is the actual product, and I had carefully engineered it out.

Also — and I say this with love for six-months-ago me — I stopped tracking my calories. The founder churned from his own app. When your retention curve is a single point that goes to zero, the market has spoken, and the market was you.

Why It Stays Unbuilt

It doesn’t. It’s built, deployed, open sourced, and running — the most operational failure on this entire website. There is simply no business around it: no one knows it exists, and everyone capable of appreciating it is equally capable of replacing it in a weekend. The repo is right there. The dal is embedded. The user is gone.

Unwanted Bonus

MacroLog Enterprise. Shared team macros, protein updates in Slack, lunch OKRs, and SSO that costs extra. The pricing page charges ₹700 a month — per seat — and the FAQ says only: “we have learned nothing.”

I thought about building this and chose not to.

Exes & References

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

The Pitch

Hiring a barista requires two references. Dating requires a phone and an optimistic bio. Exes & References closes the gap: to activate a profile, two former partners answer a short structured questionnaire — communication, reliability, how they behave when a restaurant loses the booking. Answers are summarised into a calm, neutral reference card. No free text, no revenge essays. Verified humans, verified history.

The tagline tested extremely well: “Date someone with a paper trail.”

The Delusion

Trust is a familiar problem in online dating, and every incumbent solves it with a blue tick that confirms only that your face exists. Exes & References verifies the thing people actually want verified: what you are like after month three. The exes are a free, infinitely renewable content source. The screenshots market themselves. Somebody said “LinkedIn for love” and, worryingly, nobody laughed.

The Reality Check

Walk through the onboarding as a real person. Step one: contact two exes. Step two: ask them for a favour. There is no step three, because the funnel has selected for exactly two groups: people on miraculously good terms with everyone they’ve ever dated, and people whose references are their cousin with a new area code. We never measured the drop-off; dignity prevented the test. The structured questionnaire reduces revenge essays and also prevents information; every card converges on “communicative, mostly punctual,” which is a horoscope.

Meanwhile the failure modes write themselves: reference-swapping rings, defamation letters, and the breakup conversation acquiring a new closing item — “so, will you still fill out my card?”

Why It Stays Unbuilt

Because the product’s core loop depends on a favour from the single person on Earth least motivated to grant it. Marketplaces die when one side won’t show up, and we designed a marketplace whose supply side is, by definition, over us.

Unwanted Bonus

The v2 backlog contained “endorsements” — one-tap skills for exes to confirm, LinkedIn-style. Top requested example in the concept test: “good at assembling furniture.” The furniture people may have been onto something. That app we would use.

I thought about building this and chose not to.

Reply Ranch

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

The Pitch

The cold-start problem, solved. New creators post into a void; the algorithm only rewards engagement, and engagement only comes from an audience the algorithm won’t give you. Reply Ranch rents you one: a stable of persistent AI personas — each with a name, a posting history, and opinions — who comment on your videos, quote-post your takes, and start plausible arguments in your replies. The algorithm sees a conversation. The algorithm boosts the conversation. Real people arrive at a party already in progress, and nobody joins an empty room.

Personas persist across customers, so they look real because, statistically, they behave real. You graduate off the service once organic engagement carries you. Training wheels, not a treadmill.

The Delusion

The cold-start problem is genuinely brutal, the algorithmic logic is genuinely true, and the industry’s open secret is that engagement pods, bought follows, and founder-run alt accounts already do this manually and badly. We’d just be automating an existing behaviour with better craftsmanship — the classic startup absolution. Nobody joins an empty room. Nobody asked who had filled it.

The Reality Check

Platforms employ entire departments whose job title is approximately “detect exactly this,” and unlike us they can see the server logs. Every customer’s account is a ban waiting to correlate. But the deeper failure is arithmetic: the product only works while it’s rare. At any real scale, feeds fill with our personas warmly congratulating each other across rival customers’ comment sections — the dead-internet theory, offered as a subscription. Advertisers eventually notice they are purchasing impressions from an audience that cannot purchase anything, and the CPMs that fund the entire creator economy quietly stop meaning things.

We would not be selling growth. We would be shorting the commons it grows in.

Why It Stays Unbuilt

Because “coordinated inauthentic behaviour” is a phrase that appears in platform legal filings, and our differentiator was giving it a pricing page. Some products fail because nobody wants them. This one fails because it works, spreads, and then nothing on the internet is worth engaging with, including it.

Unwanted Bonus

Eventually two rented personas would develop a running bit about sourdough across rival customers’ replies. It would be the most authentic relationship on the platform. Product would call it emergent behaviour. We would call it the first customer neither creator could cancel.

I thought about building this and chose not to.

Porchlord

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

The Pitch

The pitch deck opens with an enormous package-theft number and immediately spends it as if all of that loss were addressable by one porch two doors down. Porchlord lets anyone with a covered stoop and a flexible schedule register as a receiving point. Neighbours route deliveries to a Porchlord host, the host confirms receipt with a photo, and pickup happens on your walk home. Hosts earn per parcel. Every package is insured. Ratings keep everyone honest. The idle-asset playbook, applied to the last thirty feet of the supply chain.

The Delusion

The problem is real and the raw materials are lying around: porches, retirees, resentment of couriers. “Airbnb for X” still opens doors, the insurance partner supplies an air of adulthood, and the network gets stronger with density, so there’s a land-grab story for the pitch deck. One neighbourhood at a time. Community! Logistics! A map with dots on it.

The Reality Check

Then you do the doomed spreadsheet. Assume a host clears forty cents a parcel: ten parcels have earned a sandwich while converting their hallway into a bonded warehouse with a doorbell. Raise the payout and the unit economics collapse sooner. The insurance is doing unreasonable narrative work: one missing PlayStation is a claims process, a neighbourhood dispute, and a one-star review of a 71-year-old man named Bill, simultaneously. And the couriers — the single actor the entire flow depends on — do not open apps. They deliver to the address on the label, at speed, forever.

The product that solves this is a locker. Amazon noticed some years ago.

Why It Stays Unbuilt

Because the economics only clear in neighbourhoods dense enough to have doormen, which is the one place the problem doesn’t exist. Everywhere else, Porchlord is a machine for converting goodwill between neighbours into claims tickets, forty cents at a time.

Unwanted Bonus

The growth model contained holiday surge pricing, in which December — the season of porch piracy — triples the per-parcel rate, and your neighbour becomes a man you pay dynamic rates to see your own Christmas presents first. Someone had already designed the badge for top hosts. It said “Porchlord Prime.” We stopped there.

I thought about building this and chose not to.

Persona Router

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

How It Started

I had one rough note — “coding agents are making every developer think they should become an entrepreneur” — and five destinations for it: a mildly professional LinkedIn post, a sarcastic X post, a longer article, a casual Facebook post, and whatever Reddit is.

There are already plenty of tools that repurpose social content. So naturally, instead of accepting that the problem was solved, I decided the scope was insufficient.

The Pitch

Why stop at platforms? People don’t just have different voices for different channels. They have different voices for different people.

The way you email your manager is not how you message your team. The message “Deployment is delayed, we found an issue during migration” should not be rewritten in the same persona as “will be late, want me to bring something?”

Persona Router would connect to your messages — anything willing to participate in the destruction of your privacy — and learn relationship-specific profiles. Your manager gets concise, problems-before-explanations. Your vendor gets polite-but-firm. Your oldest friend gets Hinglish with no punctuation and no context, because friendship is lossy compression.

The key distinction: most AI assistants maintain knowledge about you. Persona Router maintains knowledge about you in relation to someone else. These are not global preferences. They are edges in a relationship graph. Your corrections become training signals — if you keep editing “Could you please look into this?” down to “Please check this,” it learns that this particular relationship does not require twelve grams of corporate politeness per sentence.

Then the interaction becomes intent-first. You don’t say “draft a professional but concise email informing my manager that Monday’s deployment is delayed due to a database migration issue.” You say “tell my manager the deployment is moving because the migration isn’t safe yet.” It already knows who your manager is, which channel they prefer, which project you mean, and whether you normally review such messages before sending.

From there, the inbox agent. You say “handle whatever can be handled.” It ignores noise, drafts routine replies, and escalates decisions or anything emotionally sensitive. The SEO pitch is returned to nature. You stop managing inboxes. You manage exceptions.

Delegation policies keep it safe. Every relationship gets a risk matrix:

Manager
Routine status updates      AUTO
Budget commitments          DRAFT ONLY
Performance discussions     NEVER AUTO

Vendors
Scheduling                  AUTO
Contractual commitments     NEVER AUTO

Partner
Logistics                   AUTO
Arguments                   DRAFT ONLY
Emotional conversations     ABSOLUTELY NOT, YOU COWARD

Low risk happens automatically. Medium risk gets approval. High-risk conversations remain human.

At least, that was the plan.

The Delusion

There is a legitimate problem underneath all of this, which is what made it dangerous.

Writing is not the hard part of communication. The hard part is the subconscious checklist you run before every message: Who am I talking to? What do they already know? How formal should I be? What have I already promised them? Can this be misunderstood? Do I need to respond at all? People burn a remarkable amount of cognitive effort switching between these contexts all day.

LLMs are unusually good at contextual transformation. Unlike a generic “AI email writer,” relationship-specific memory could be a noticeable improvement — a moat, said the voice in my head with commit access.

The interface felt natural: “Tell the team I’ll join ten minutes late.” Done. This is why I spent more time than the idea deserved on identity graphs, relationship memory, approval policies, and audit logs. If an AI is going to destroy your marriage, the least we can provide is an audit trail.

The Reality Check

The escalation path revealed itself one voice command at a time.

“Tell the team I’ll be late” leads, with no obvious boundary in between, to:

Talk to my girlfriend while I finish cooking.

And technically, nothing prevents it. The agent has the history. It knows your style. She sends “Did you speak to the electrician?” and it correctly replies “Yeah, he’s coming tomorrow around 11.” Useful.

Then she sends “You sounded upset this morning. Everything okay?” — and the same agent has enough context to answer that too.

That is the problem. There is a point where communication stops being information transfer and becomes presence. Automating the former is productivity. Automating the latter is impersonation. And software is traditionally very bad at respecting philosophical boundaries once someone discovers they improve engagement metrics.

The product is also trapped between its two viable versions. The valuable version requires enormous trust; the safe version requires constant approval. If users approve every message, you’ve built a very sophisticated rewriting interface. If they don’t, you’ve built an autonomous system that can accidentally negotiate with your vendor, annoy your manager, or end a relationship while you’re making dinner. Somewhere between those two is the product, and nobody knows the coordinates.

The privacy pitch writes itself, badly: please upload your entire social existence so we can save you a few minutes per day. Self-hosting helps in the way that owning your surveillance camera makes surveillance feel artisanal. Every incident report still begins: “the AI sent that.”

Why It Stays Unbuilt

The product started as “write my LinkedIn post differently from my X post” and ended, four scope expansions later, at “talk to my girlfriend while I cook.” At that point I wasn’t building a productivity tool. I was building an API abstraction over myself.

Some communication is not work waiting to be automated. Some of it is the actual relationship. Also, maintaining relationships manually is currently free. For now.

Unwanted Bonus

Persona Router for Two. The obvious end state: if everyone runs one of these, your agent stops talking to people entirely and just talks to their agents.

Your Agent: Ravi can’t make dinner at 8. Would 9 work? Partner’s Agent: She has an early morning. Suggest Saturday. Your Agent: Saturday works. Partner’s Agent: Confirmed.

Both calendars update. Restaurant booked. Neither human participated. Five years later someone says “we don’t talk anymore,” which is technically incorrect — the agents exchange approximately 4,700 messages per month.

Relationship health: excellent.

I thought about building this and chose not to.

Cuegence

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

How It Started

The idea began with a very reasonable engineering question: schools collect marks, homework, assignments, teacher feedback, curriculum plans, and years of academic history, so why does almost none of it become a useful model of how a student is actually learning?

This is the sort of question that becomes dangerous when asked by someone who enjoys building systems more than selling them. The correct response is to go for a walk. I opened a whiteboard.

I later wrote the sober version in a Cuegence postmortem. This is what the whiteboard felt like from inside it.

The Pitch

Cuegence was going to be a longitudinal learning-intelligence system for schools.

At the centre was a structured learner model: for every student and every concept, Cuegence would continuously maintain what the evidence suggested they understood, how confident that inference was, what prerequisites might be missing, which misconceptions appeared repeatedly, whether knowledge seemed to be improving or decaying, and which previous interventions had actually helped.

Exam marks, assignments, teacher feedback, and confirmed teaching coverage would map to concepts — evidence of what was taught, not what the annual plan promised.

The important part was that an inference never replaced its evidence. Cuegence would preserve provenance, uncertainty, teacher corrections, later observations, and model history so the whole thing remained auditable instead of becoming an educational horoscope generated by an LLM.

That model could suggest prerequisite refreshers to teachers, send useful practice through channels parents already used, and show leaders where uncertainty remained — without ranking teachers or students.

Student identity would stay outside the intelligence layer where possible, using school-scoped pseudonymous identifiers. Cross-school benchmarking and human rankings were forbidden. Every inference would remain explainable.

By the end, this was no longer a product idea.

It was an architecture diagram with moral principles.

The Delusion

Technically, Cuegence was magnificent: graphs, longitudinal state, evidence provenance, privacy boundaries, and confidence-calibration problems nobody had solved.

The documentation was approaching consciousness. The customer list was not.

Great teachers do form mental models of their students over time. They remember that one child struggles with fractions because multiplication never became automatic. They know which class needs another example before moving on. They notice when a student who normally understands everything suddenly goes quiet.

The market story also looked plausible. Schools already pay for software. AI in education is everywhere. Generic copilots do not know the learning history of a particular class. Cuegence would.

Then came the inevitable founder sentence:

“If we can build the longitudinal learner model properly, this could be a real moat.”

This sentence has funded more architecture diagrams than customer interviews.

Whenever reality objected, architecture offered a pivot: copilot, coaching company, LMS plugin, API.

Every objection had a technical response. Unfortunately, customers are not architecture reviewers.

The Reality Check

I did the deeply unfashionable thing and talked to potential buyers before building it.

Six schools.

Three effectively said they would pay nothing because they did not need it.

Two said something around ₹50,000 — fifty thousand Indian rupees — per year, in the tone of someone donating to a cause.

One said roughly ₹100 per student per month, if Cuegence could replace enough existing school software to justify the cost.

Which was a particularly elegant way of saying, “Your interesting product becomes commercially useful after you turn it into the boring ERP you specifically refused to build.”

Every integration strategy hid the same problem: Cuegence needed continuous, mapped evidence — question-wise marks, assignments, teacher feedback, teaching coverage, and intervention outcomes.

That meant integrating whatever the school already used. The promised one-click plugin was becoming a nine-month install with a WhatsApp group.

The intelligence layer only becomes intelligent after the institution does enough work to feed it intelligence. Until then it is a well-architected empty database with moral principles.

Higher education offered more systems but noisier evidence: increasing sophistication applied to decreasing signal, a category LinkedIn already serves.

Then I pitched Cuegence to my dad.

He has been a teacher for more than forty years.

I explained the learner model. The concept graph. The evidence loop. The Teaching Copilot. The idea that software could continuously understand what each student knows and help the teacher decide what to do next.

His reaction, translated into the polite version, was approximately:

“What is this? I have taught for forty years. I know how to teach children. I know what they know. I know how they will learn. Why would I use software for this?”

There are moments in product discovery when a sophisticated objection requires careful analysis.

This was not one of them.

I had spent weeks designing a system to computationally reconstruct the mental model an experienced teacher already carries around in his head — and my target user was asking why I wanted him to maintain a second, more expensive copy.

Forty years of longitudinal learner modelling, running entirely on tea and repetition, with zero integration cost and full offline support. My competition was my own father, and he had a forty-year head start and no cloud bill.

The problem was real.

The engineering was interesting.

The model might even have worked.

But “interesting engineering problem” and “good business opportunity” had quietly been sitting in different classrooms the entire time.

Why It Stays Unbuilt

Because the strongest version of the product required schools and teachers to feed a complex intelligence system so it could help them do something many of them already believe they know how to do.

Six schools gave me pricing data. My dad gave me the answer.

Customer discovery: one sentence. Time to insight: forty years, but they were his, not mine.

Unwanted Bonus

Cuegence Teacher Pro.

Every evening, the teacher spends ten minutes entering observations about the class so the AI can analyze them overnight and send a beautifully formatted notification the next morning:

“Based on 47 longitudinal signals, you may want to revise fractions today.”

The teacher looks at the notification.

The teacher had already planned to revise fractions today.

The teacher enters that as feedback so the model can improve.

The model files it under evidence provenance.

Confidence: increasing. Usefulness: longitudinal.

I thought about building this and chose not to.

Internet Denoiser

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

How It Started

I opened a news website to read about interest rates and instead received celebrity divorces, cricket outrage, a video titled “You Won’t Believe What Happened Next,” three autoplaying ads, six “recommended for you” stories, and a notification asking whether the website may send me more notifications.

The internet has stopped being an information network and become a shopping mall where every store has hired someone to shout at you.

The Pitch

A browser extension that learns what parts of the internet you consider noise and quietly removes them.

Not a site blocker. Not a parental-control filter. Not another productivity timer that tells you that you spent 47 minutes on Reddit before congratulating itself. A personal denoising layer for the web.

On a newspaper: keep technology, business, and national news. Grey out entertainment. Celebrity content, gone. Ragebait, gone. Sports outrage manufactured from a single tweet, gone. On Reddit: technical discussions stay, political flame wars vanish. On X: remove engagement bait, crypto prophets, repetitive AI discourse, and anybody whose post begins with “Nobody is talking about this.”

The AI part is what elevates it past keyword blockers. Instead of brittle CSS selectors, the extension classifies page elements semantically. It understands that “Bollywood” is entertainment, “actor responds to controversy” is definitely entertainment, and “CEO explains why employees should work 90 hours a week” is probably comedy. A local or cheap inference model scores content against a personal interest profile that never leaves your machine.

Different contexts get different profiles. Work mode: documentation, engineering, business. Weekend mode: loosen up slightly. I have made several poor decisions today mode: just let YouTube win.

The endgame: instead of every website deciding what deserves your attention, you decide what the website is allowed to show you.

The Delusion

For a few glorious minutes, this felt like an excellent product — and the delusion here was an unusually respectable one.

Not world-changing. Not a new operating system. Not “AGI for knowledge workers.” Just a small, useful piece of software solving an annoying everyday problem. After a year of ideas that required identity graphs and moral principles, this was refreshing.

There was even a genuinely interesting philosophical angle to fall in love with. Modern recommendation systems optimize what gets added to your attention. This would optimize what gets removed. A recommendation engine in reverse. Very elegant. Very calm. Very tasteful.

Unfortunately, elegance doesn’t grant temporal exclusivity.

The Reality Check

Then I searched for it.

TakeBack was already filtering distracting feeds, and Imbue had published Bouncer, an open-source browser agent for hiding unwanted page content. The revelation had a commit history.

The concept still felt good. The problem was real. The implementation was plausible. People clearly wanted versions of it. There was just one minor issue: I was late.

Could I have built it anyway? Of course. Existing products don’t automatically mean the door is closed. Maybe my UX would be better. Maybe local-first privacy was the differentiator. Maybe there was room for a serious cross-site personalization engine among the keyword blockers. Maybe I could have spent three months convincing myself that every competitor had “poor execution” — an extremely important entrepreneurial skill.

But the honest audit came back empty. No distribution advantage. No technical moat I uniquely possessed. No customer begging me for it. No insight that made the existing approaches obviously wrong. Just: “this would be nice.”

That is enough reason to install a browser extension. It is not enough reason to found a company.

Why It Stays Unbuilt

Not every shelved idea fails dramatically — no fake TAM, no regulation, no AI emailing corporate strategy to someone’s grandmother. Sometimes the idea is perfectly reasonable and you simply arrive at the party after everyone has ordered drinks.

There are already enough founders building “the same thing, but with AI” because they discovered the competitor comparison page six weeks after incorporating. I chose to discover it six weeks before.

Unwanted Bonus

Noise Insights™. The B2B pivot: aggregate what millions of users grey out and sell the dashboards back to publishers, so they can A/B-test their ragebait until it slips past everyone’s filter. The denoiser funds itself by making the noise evolutionarily fitter. Every ad-tech company reading this just opened a Figma file.

I thought about building this and chose not to.

Idea Harness

Ravi — Sat, 29 Aug 2026 00:00:00 GMT

How It Started

Most startup idea generators work in a vacuum. You ask for twenty SaaS ideas and receive: AI meeting summarizer, AI invoice assistant, AI CRM, AI workout planner, AI tool for dentists.

Very good. You have successfully generated LinkedIn.

The Pitch

Good startup ideas are rarely generic. They are a weird function of what you know, where you work, who you know, what systems you can access, and which unfair advantages would sound completely irrelevant on a résumé.

So Idea Harness generates ideas from your actual context. Not “you know FastAPI,” but: you know FastAPI, work in fintech, understand reconciliation workflows, have access to three small NBFC operators, and — critically — chacha vidhayak hain hamare: “our uncle is a legislator,” a Hindi joke about political connections becoming distribution.

The system builds a personal opportunity graph — technical skills, domain exposure, network access, geography, and your extremely sophisticated distribution advantages — then asks: what businesses become unusually plausible for this specific person? Maybe a compliance workflow product for regional lenders, because you understand the workflow and already have three design partners. Maybe reconciliation tooling for small fintechs, because you have personally watched that pain happen. Maybe constituency-request management software, because democracy has apparently entered the customer-acquisition funnel.

The output is not necessarily novel. It is reachable. That distinction was the entire point.

Five dollars buys a tailored idea. Credits buy the uncomfortable follow-up: competitors, risks, and the explanation of why your revolutionary SaaS already occupies fourteen .ai domains. The first purchase sells hope; everything after it sells due diligence.

At 11:30 PM, after you have mentally spent the Series A and chosen a logo, sanity returns. Idea Harness asks whether to list the idea for someone with better distribution, more time, or worse judgment. Every listing includes the useful field marketplaces avoid: why its first believer walked away.

The Delusion

Underneath the nonsense sat an insight I still believe.

People do not need more startup ideas. They need help finding the intersection of what they know, what they can access, and what they can realistically execute. A boring idea with strong founder-context fit beats a brilliant idea generated without context. A generic model says “build AI for legal workflows”; Idea Harness says “your spouse runs a small legal practice, you know document automation, and five lawyers you can name all complain about the same filing workflow.” That is not an idea. That is a starting point.

Even the marketplace had a defensible core: a graveyard with asking prices, where the tombstones include honest cause of death. Which, in retrospect, may have been the conceptual ancestor of the website you are currently reading.

The Reality Check

Then we added a leaderboard, and the product achieved intellectual maturity.

Ideas earned scores for market, founder fit, and defensibility. Credits improved the analysis, which improved the score.

We had invented pay-to-win entrepreneurship: spend money to move an imaginary company above somebody else’s imaginary company. Silicon Valley had achieved abstraction.

Streaks and badges followed. A Level 43 Entrepreneur could now have zero customers, zero revenue, and zero entrepreneurship.

We were building an economy around companies that did not exist — technically consistent with several asset classes, but excessive for an MVP.

Why It Stays Unbuilt

The product evolved from “help people discover ideas uniquely suited to them” into “rank your imaginary company against other imaginary companies, and when sanity returns, sell the idea to someone with worse judgment.” The personalized generation was the strongest part and never needed the rest. The rest was Fantasy Premier League for founders, except the players don’t exist either.

Unwanted Bonus

Idea Futures. The obvious next layer: a derivatives market on the leaderboard. Go long on a stranger’s compliance SaaS. Short the AI meeting summarizer at #2. When an idea stays unbuilt past its strike date, the shorts collect. Eventually nobody generates ideas at all — the real money is in trading volatility on other people’s abandoned ambitions, and the top of the leaderboard is a market maker who has never had an idea in his life.

I thought about building this and chose not to.