Pre-seed brief
Investor & founding partner
August 2026
The reels you saved, turned into things you actually do.
Synapse listens to every reel you save, writes down what it actually said, and hands it back at the moment you can use it — the recipe at the stove, the restaurant at the corner, the deadline a week before it passes.
200B
Reels played per day across Instagram and Facebook — double late 2024.
Zuckerberg, Meta Q3 2025 call
3.0B
Instagram monthly actives. India is the largest single market at 480M.
Meta Q3 2025 · Statista
40%
Of 18–24s go to Instagram or TikTok, not Google, to find somewhere to eat.
Google internal research, 2022
0
Apps that bring a saved reel back to you at the moment it becomes useful.
See slide 15
00:45The problem
Saving is not remembering. It only feels like it.
Tapping the bookmark gives you the satisfaction of having dealt with something. Nothing was dealt with. The reel went into a grid with no titles, no text, no search that works, and nothing that ever brings it back.
There is no public number for how many saved reels ever get opened again — because Instagram doesn't give you one. That absence is the product opportunity.
Seventy saves. One of them is the restaurant you are standing outside of right now. You will not find it, because you no longer remember it exists.
2 taps
deep, unlabelled, behind a hamburger menu — where Saved lives inside Instagram.
88%
of interviewees in a UX study could not recall how to reach their saved posts.Small-sample study — flagged as low confidence in the appendix.
~50%
of all time on Instagram is now Reels — so most saves are video, the format with no text to search.
177
videos the average Instagram user watches per day. The good ones vanish into the same grid.
01:40Diagnosis
It isn't that people are disorganised. Three things are structurally missing.
Missing · one
A text layer
Everything useful in a reel is spoken or burned into the frame. The caption is marketing. So there is nothing to index, nothing to search, nothing to skim — the information exists only as pixels and sound.
Consequence
You must rewatch a 45-second video to recover eight seconds of value.
Missing · two
A way back in
Retrieval is entirely on you: remember that you saved something, remember roughly what it was, then hunt a wall of thumbnails. Collections exist, but they ask you to stop scrolling and file things — nobody does that mid-scroll.
Consequence
The library grows faster than any human can index it.
Missing · three — the big one
A trigger
A saved restaurant is useful when you are near it. A saved recipe is useful at 7pm on a Tuesday. A saved event is useful the week before. Nothing in any of these apps knows where you are or what day it is.
Consequence
Intent expires silently. This is the gap we build in.
Read-it-later has already failed once at exactly this point. Mozilla shut Pocket down on 8 July 2025 and deleted the archives on 8 October, saying browsing habits had changed. Pocket stored things beautifully. It never brought them back. Storage was never the product.
02:30Who feels it
Four kinds of saver. All four are saving with intent.
Saves are the highest-intent action on Instagram — a save says "I mean to do this", not "I liked this". Socialinsider's read of 35 million posts across 447,613 pages puts food, drink and travel at the top of the save table. Those are exactly the categories where intent has a place, a price and a date attached.
We are not building for people who save memes. We are building for people who save plans.
The cook
largest, most provenSaves recipes she never makes because the video is a bad instruction manual. Already spending money here — see slide 15.
The eater-out
our wedgeSaves cafés, dhabas, bars, street food. Has 60 places saved in his own city and eats at the same four. Nothing on earth reminds him.
The trip planner
Saves 40 reels before a trip, then rebuilds it all by hand in Notes and Google Maps the night before flying.
The professional
pays 3× moreCreators, social managers, chefs, trainers. Saves competitors to study hooks and formats. Needs transcripts and patterns, not vibes.
03:20The fix
One share. The video becomes something you can use.
Synapse watches the reel so you never have to again. It listens to what was actually said, writes it down, works out what kind of thing it is, and files it where you will trip over it at the right moment.
POV: best kebab in Old Delhi and nobody talks about it 🔥 save this before the crowd finds out
47 seconds. No address,
no name in the text.
Qureshi Kebab Corner
Gali Kababian, Jama Masjid, Old Delhi
4 other saves within 600 m. Synapse will tell you when you're next in Chandni Chowk.
04:15How it works
Five steps, and four of them are invisible.
The user does one thing: share. Everything after that is ours.
01 · Capture
Share sheet, or bulk import
Share a reel from Instagram, Facebook, TikTok or YouTube Shorts. Or import your whole saved history in one go from your Instagram data export.
iOS Share Extension · Android Intent
02 · Hear
Speech and on-screen text
Audio is transcribed; burned-in captions are read with OCR. Hindi, Hinglish and regional code-switching handled, because that is what Indian reels sound like.
Whisper-class ASR · frame OCR
03 · Understand
Structure, not summary
One LLM pass returns typed JSON: category, a two-line explanation, ordered action items, and entities — places, dates, prices, products, people.
Small-model structured output
04 · File
Resolved and indexed
Place names are resolved to real coordinates. Dates become calendar events. Everything is embedded so it can be found by meaning, not by keyword.
Places API · pgvector
05 · Surface
Brought back at the moment
A checklist when you're cooking. A nudge when you're 200 m away. A reminder the week before the event. This is the step nobody else does.
Geofence · scheduled push
Retention policy by design: we keep the transcript, the structured record and a thumbnail. We do not keep the video file. This is a deliberate product, cost and legal decision — see slide 13.
05:10Capability · Focus mode
A 45-second video is a terrible set of instructions.
You cannot scrub a timeline with dough on your hands. Focus mode throws the video away and gives you the thing the video was trying to be: a list, at arm's length, one step at a time.
- Ingredients first, quantities normalised, exportable to a shopping list.
- Tap anywhere to advance. Big type, no precision required.
- Jump back to the exact second of the video for any step you don't believe.
This is the feature people will screenshot. It is also the cheapest one we build — it is pure output from the pipeline, with no new permissions and no native device work.
Chettinad chicken
Step 3 of 6 · 40 min · serves 4
05:50Capability · Places
You already built a list of places — inside a video app.
Every saved reel with a location becomes a pin. Forty scattered saves become one map of your own city, and one map per trip. Then the app does the part a list can't: it speaks up when you are standing next to something you wanted.
- Arrive-and-remind. A single quiet notification when you're within 200 m of a place you saved.
- Trip mode. Land in a new city and the pins you saved months ago cluster around you.
- What to order travels with the pin — the actual line from the transcript, not a star rating.
Geofencing is the one feature that needs real native work and a background-location permission. We ship it in month 4, after the pipeline has earned the trust to ask.
Synapse · now
Qureshi Kebab Corner is 180 m away
You saved this in March. Order the buffalo galouti, and go before 8 pm.
06:30Capability · Recall
Search the way you actually remember things.
Nobody remembers the creator's handle. People remember a detail: dark wood, no eggs, the highway. Because we hold the transcript and not just the caption, a half-memory is enough to find it.
Instagram search cannot do this. It has no text layer for spoken words inside your saves, so it can only match captions and hashtags the creator happened to write.
Match scores are cosine similarity on the transcript embedding. The reel that wins here never had the word "dhaba" in its caption — it was spoken at second 12.
07:05Capability · Dates
Save it in March. Get told about it in June.
Reels are full of dates that expire quietly: ticket drops, pop-ups, exhibition closings, application deadlines, sale windows. A save is the worst possible place to keep a deadline. Synapse pulls the date out and puts it where dates belong.
- Parsed, not guessed. "Only till the 21st" resolves against the day the reel was posted.
- Your calendar, not ours. One tap writes to the device calendar with the link and the transcript in the notes.
- Two nudges — one a week out, one the morning of. Then it gets out of the way.
Kala Ghoda Arts Festival — last weekend
The creator says "it wraps up on the 9th". Posted 14 January, so: Sunday 9 February, Rampart Row, Fort, Mumbai. Free entry.
Synapse · Sunday, 8:10 am
Today is the last day of Kala Ghoda
You saved three reels about it. Two are food stalls on the same street — open the map.
07:40Capability · Shared
Saving is usually a two-person activity. Every app treats it as one.
Couples planning a trip, flatmates planning dinner, a content team studying competitors — they already forward reels to each other in DMs, where the reels die a second death. A shared folder in Synapse turns that stream into one document.
"Vietnam · Feb"
Two people, 46 forwarded reels, one itinerary — day-ordered, mapped, with everything both of you said you wanted to eat. Nobody has to be the planner.
"Things we'll actually cook"
Recipes merge into one weekly plan and one grocery list. The list is the product; the reels were just the raw material.
"Competitor teardown, Q3"
Drop 200 reels from a niche and get the recurring hooks, formats, opening lines and lengths, with the transcript evidence attached. This is the tier that pays three times what a consumer does.
Sharing is also our cheapest growth loop: an invited collaborator arrives with intent, sees a folder that is already valuable, and has a reason to install before they have a reason to pay. Every shared checklist and itinerary also has a public web link with a footer.
08:20The outcome
What actually changes, stated as numbers we will be judged on.
| What we're changing | Today | Design target |
|---|---|---|
| Searchable text per saved reel | ~0 words caption only, if any | ~110 spoken words + on-screen text |
| Finding one specific save | Scroll an untitled grid | Under 2 seconds, by meaning |
| Cooking from a saved reel | Rewatch 45 s, scrub 4–6× | 6 steps, tap to advance |
| Reaching a saved restaurant | No mechanism exists | Nudge within 200 m |
| Catching a saved deadline | Expires silently | 2 reminders + calendar entry |
North-star metric
Saves acted on
The share of a user's library that produced a real action inside 30 days — a place visited, a recipe cooked, an event attended, a checklist completed.
Baseline today
≈ 0%
no mechanism exists
Target, month 6
12%
measured, published monthly
These are targets, not results. We have no users yet. The point of stating them this precisely is that they are all falsifiable inside six months — and every one of them is measurable from our own event stream, not a survey.
09:10Feasibility · the honest bit
The obvious version is impossible. This one isn't.
This is the first question a serious investor asks, so it is slide 13 and not an appendix.
Q1 Can you just OAuth into Instagram and pull a user's saved reels?
No, and anyone who says otherwise hasn't read the docs. The Instagram API with Instagram Login exposes four scopes — instagram_business_basic, _content_publish, _manage_messages, _manage_comments — and only for Business and Creator accounts. There is no saved-media scope and no Collections endpoint. The private internal routes that third-party tools use are already returning 404 with the server message "Dead endpoint … unshipped". We do not build on those.
Q2 Then how does anything get in?
Three legitimate paths, all user-initiated. (1) The share sheet — an iOS Share Extension and an Android intent, one tap from any reel, and it covers Instagram, Facebook, TikTok and Shorts equally. (2) Bulk history import — the user requests their own Instagram data export and uploads it to us, which kills the empty-app problem on day one. We validate that file field-by-field in week one and will state publicly what it does and does not contain. (3) Paste a link.
Q3 Are you storing other people's videos?
No. We keep the transcript, the structured record, a thumbnail and a deep link back to the original. Audio is discarded once transcribed; the video file is never persisted. That is simultaneously the cheapest, the most defensible and the most honest design — and it is the same posture that makes a creator opt-out trivial to honour.
Q4 Doesn't this take value from creators?
It sends traffic back. Every step deep-links to the exact second of the original reel, and the creator's handle travels with the record. A saved reel that gets rewatched at the stove is worth more to a creator than one that is never opened again. We publish an opt-out register before launch.
Q5 Will transcription survive an Indian reel with background music and Hinglish?
Not out of the box, and we've priced that in. Whisper large-v3 sits around 9–14% word error rate on Hindi and roughly 18% on clear Hindi speech, degrading past 30% with strong regional accents; code-switched speech raises WER by a further 30–50%. So Hinglish gets a fine-tuned model, and on-screen-text OCR runs as an independent second signal. This is a line item in the plan, not an assumption.
10:10Market · built bottom-up
No top-down hand-waving. Every layer shows its arithmetic.
3.0B
people on Instagram every month, before adding Facebook, TikTok and Shorts
Meta Q3 2025. Reels alone is played 200B times a day and runs at a $50B+ annual rate.
165M
intent savers in our two launch markets
India 480M + US 181M Instagram users = 661M · × 25% who keep more than 20 saves carrying a place, recipe or date = 165M.
25% is an assumption, not a finding — see below
$4.9M
ARR at year five, on deliberately unheroic conversion
165M × 2.5% install = 4.1M installs · × 3.5% paid = 145K subscribers · × $34 net revenue per subscriber per year = $4.9M.
Every percentage in those three lines is a number you can argue with, and we would rather you did it here than after the cheque. A market slide nobody can attack is a market slide nobody believes.
The upside case, stated plainly
The SOM above assumes we convert 3.5% of installs and reach 2.5% of intent savers. If Synapse instead reaches 1% of the 165M intent-saver base as paying subscribers — 1.65 million people — the same $34 net ARPPU gives $56M ARR. Cal AI reached a reported $30M+ ARR inside two years on a single-purpose consumer AI utility, bootstrapped, and sold to MyFitnessPal in December 2025. The ceiling is not the problem.
The one number we are missing
Nobody publishes how many people use Instagram Saves or how much they keep in there — Meta doesn't disclose it. So the 25% is our assumption and we treat it as the single biggest risk in the model. We settle it two ways in the first eight weeks: a 500-person survey across both markets, and then the real thing — a histogram of actual save-counts from the first thousand bulk imports. That measurement is free, it is a by-product of the product working, and it is worth more than any report we could buy.
11:15Competition
Parts of this exist. The last two columns don't.
| Product | What it is | Price | Scale | Ingests reels | Reads the audio | Action items | Brings it back by place & date |
|---|---|---|---|---|---|---|---|
| Instagram / Facebook Saves | The default. A grid. | Free | 3.0B MAU | native | — | — | — |
| ReciMe | Recipes out of reels | $59.99/yr | 10M users claimed | ✓ | ✓ | recipes | — |
| Wanderlog | Trip itineraries | $39.99/yr | Top-5 trip planner | links | — | manual | no nudge |
| mymind | Visual AI bookmarking | $13/mo | Niche, premium | images | — | — | — |
| Readwise Reader | Read-it-later, for text | $9.99/mo | Prosumer | — | YouTube | — | — |
| Raindrop | Bookmark manager | $3/mo | Broad, cheap | links | — | — | — |
| Google Maps Lists | Saved places | Free | Universal | — | — | — | manual |
| Read-it-later, 2007–2025 | Dead | Shut 8 Jul 2025 | — | — | — | — | |
| Synapse | Every category, then the nudge | $4.99/mo | Pre-launch | ✓ | ✓ | ✓ | ✓ |
ReciMe is the most useful fact on this slide. It claims 10 million users and charges $59.99 a year to do one category of what we're describing — proof that people will pay real money to get information out of a reel, and proof that the pipeline is buildable. It also tells us not to lead with recipes.
So we lead with places. Nobody has connected a saved reel to where you are standing. That column is empty across the entire table, it is the column that turns a library into a habit, and it is the one an incumbent is least likely to bother with — Instagram's incentive is time inside Instagram, not sending you out of the door.
12:15Unit economics
Processing one reel costs a third of a US cent.
Costed on published list prices in August 2026, for a 45-second reel — the median length that performs on Reels. No volume discounts assumed, no self-hosting assumed.
| Step | List price | Per reel |
|---|---|---|
| Transcription — 45 s of audio | $0.003 / min gpt-4o-mini-transcribe | $0.00225 |
| Structured extraction — 900 in / 450 out tokens | $0.10 / $0.40 per 1M Gemini 2.5 Flash-Lite | $0.00027 |
| Place resolution — 30% of reels, 60% cache hit | $5.00 / 1,000 Places Essentials | $0.00060 |
| Embedding for search — 200 tokens | $0.02 / 1M | $0.00001 |
| Storage — transcript + thumbnail, 42 KB | $0.015 / GB-mo | $0.00001 |
| Total per reel | — | $0.0032 |
Deepgram Nova-3 batch at $0.0043/min is the failover transcription route and lands the total at $0.0042 — a 30% cost increase that changes nothing about the margin story. That is the definition of a robust cost base.
Contribution margin per Pro subscriber, monthly
Where a $4.99 month actually goes:
- $4.05Contribution — 81% of the sticker price, 95% of net revenue
- $0.75Store commission at 15%
- $0.19All AI and infrastructure, at 60 reels a month
A heavy user at 250 reels a month costs $0.80 and still contributes 81% of net. There is no plausible usage pattern that breaks this.
The cost that actually needs managing
Importing a new user's 300 historical saves costs $0.96 before they have paid us anything. That single number is why the free tier caps the bulk import at 30 items and why full-history import is a Pro unlock. It is a pricing decision derived from a cost, not a growth-hack.
LTV against CAC
At 12% monthly churn — slightly better than RevenueCat's 13–14% median — a monthly Pro subscriber lives 8.3 months: $41 gross, $32 contribution. Against a target blended CAC of $14 per paying user that is 2.3×, and the benchmark range for CAC per paying user is $20–80. We get under it only by keeping most installs organic, which is why go-to-market is content and share-links rather than paid ads.
13:10Business model
Free proves the magic. Pro sells the memory.
Free
$0
20 reels a month. 30-item history import. Summaries, action items, search.
No map. No nudges. The free tier proves the pipeline works; the paid tier is what remembers for you.
Costs us ≤ $0.16/user/month
Pro — the core
95% margin$4.99 /mo
$39/yr · India ₹399/mo, ₹1,999/yr
Unlimited saves, full history import, the map, arrive-and-remind nudges, calendar sync, Focus mode, exports to Notion, Reminders and Calendar.
Annual-first pricing — cheap annual plans retain up to 36% at 12 months versus 6.7% on expensive monthlies
Teams
3× ARPU$16 /seat/mo
Shared brains, trend synthesis across a folder of competitor reels, transcript export, seat management.
Sold to creators, social teams and agencies — the segment that treats this as a research tool and expenses it.
Two release valves
$5 / 100
Credits for people who refuse subscriptions. One reel = one credit. Protects margin at any volume.
$99
Founding member, capped at 1,000 seats. Funds the first year of infrastructure and buys a cohort that is financially invested in telling us the truth.
Why two currencies and not one. India generates about $0.03 per download against $0.20+ in Southeast Asia and Latin America. Pricing India at the US number would be a rounding error in revenue and a wall in adoption — so India gets a local price and becomes our proving ground and our talent base, while the US, UK, EU and SEA carry the ARPU.
And why that is not a consolation prize. India's app spending hit a record $345M in Q2 2026, up 35% year on year, and non-gaming categories are now 68% of Indian app revenue against 58% three years ago. Revenue per download has more than doubled in three and a half years. We are pricing for where that curve is going, not where it has been.
14:00The model
Three years, with every assumption on the table.
| Year 1 | Year 2 | Year 3 | |
|---|---|---|---|
| Cumulative installs | 75K | 500K | 2.0M |
| Free → paid conversion | 2.5% | 3.2% | 3.5% |
| Paying consumers | 1,875 | 16,000 | 70,000 |
| Team seats | — | 300 | 2,000 |
| Net revenue per subscriber / yr | $33.50 | $33.50 | $33.50 |
| Founding-member sales | 400 × $99 | — | — |
| Revenue | $96K | $585K | $2.67M |
| Gross margin | 88% | 91% | 92% |
Where each assumption comes from
- 2.5% conversion in year one — RevenueCat's median for freemium apps is 2.2%. We start marginally above it and improve with the paywall, not with hope. A hard paywall would convert 12.1%, and we are deliberately not using one: the free tier is the growth engine.
- $33.50 net per subscriber per year — a 60/40 annual-to-monthly mix at $39 and $4.99, less the 15% small-business store commission, less AI cost.
- 12% monthly churn — against a 13–14% median. Our bet is that a personal library gets harder to leave every month, which is the opposite of how a calorie counter behaves.
- 15% store fee — Apple's Small Business Program under $1M in proceeds, and Google's automatic 15% on the first $1M with subscriptions at 15% from day one. We model 30% from year four.
What this model is not
It is not a forecast. It is an arithmetic consequence of four assumptions, and the whole point of writing it this way is that you can attack any one of them and watch the answer move. Year one revenue is small and deliberately so — year one buys the retention data that determines whether years two and three are worth funding at all.
15:00Go to market
The first thousand come from video, because the problem is video.
A product that turns reels into checklists demos itself in a nine-second reel. That is the whole strategy: the ad and the product are the same artefact.
| Channel | Target | The arithmetic, and why we believe it |
|---|---|---|
| Short-form demo content Reels, Shorts, TikTok | 600 installs | Split-screen: someone pausing a cooking reel six times on the left, our checklist on the right. At ~110K cumulative views → 3% to the landing page → 30% install. One post that works does this alone; we plan for eight that don't. |
| Build-in-public tutorials long-form, existing channel | 150 installs | Episodes on the actual pipeline — share extension, Whisper, structured extraction. Attracts developers who forgive rough edges and file good bug reports. Slow channel, high-quality cohort, and it costs us nothing we weren't doing anyway. |
| Wedge communities city and travel subreddits, food groups | 150 installs | Lead with places, not recipes: "here are the 60 cafés you saved in Bangalore, on one map." Post where people already complain about losing saves. No spam, one thread per community, founder-signed. |
| Product Hunt + Show HN one day, once | 100 installs | Framed on the honest achievement — a text layer over video you already saved — with the founding-member deal as the launch-day perk. We will not frame it as getting around Meta's API. That headline invites exactly the attention we don't want, and it isn't what we built. |
| Shared links the compounding one | Loop | Every checklist, itinerary and map is a public web page with a footer. The recipient arrives already looking at something useful. Target: 0.35 installs per paying user per month by month nine — which is what turns paid CAC from a requirement into an option. |
Paid acquisition is a measuring instrument, not a growth plan. We will spend a capped $3K across iOS and Android purely to learn our real cost per paying user — benchmark cost per install is $4.70 on iOS and $3.70 on Android, and CAC per paying user runs $20–80. We do not scale paid until contribution per subscriber is proven at 2.5× or better.
Sequence matters more than volume. Recipes is the biggest wedge and also the one ReciMe already owns with a claimed 10 million users. So places goes first — uncontested, and it is where the nudge lives. Recipes becomes the second wave, entered with a feature nobody there has: the reminder that you have the ingredients and it is Tuesday.
16:00Defensibility
The pipeline is copyable in a weekend. These three things are not.
One · switching cost that grows
Your own library
Month one, Synapse holds 40 records. Month twelve it holds 900, indexed, mapped, and tied to places you actually went. Leaving means going back to a grid of thumbnails. This is the opposite dynamic to a calorie tracker, where every month makes you more likely to cancel.
Two · nobody's tenant
Cross-platform by design
Because ingestion is the share sheet rather than one company's API, Instagram, Facebook, TikTok and Shorts are all equal inputs and none of them can switch us off. A single-platform version of this product would be one policy change from dead. Ours is one share sheet from every platform that will ever exist.
Three · the asset nobody has
Structured save-intent
A save is the strongest unmonetised intent signal on the internet: not "I liked this", but "I mean to do this". Instagram has the saves and no structure. Google has the searches and no saves. We would hold, in aggregate, a ranked picture of what people in a city actually intend to try next.
The limit on that third point, said out loud. It is a year-three asset and it is worth nothing at launch. It only becomes real at hundreds of thousands of users, and it is worth stating plainly what we will not do with it: we will not sell personal data, and we will not resell individual saves. The value is in ranking and recommendation inside the product — knowing that 4,000 people in Mumbai saved a place this month is a better restaurant signal than any review site has, and it belongs to the users who generated it.
What that leaves as the real defence in year one: being the only app that fires at the right moment, and being noticeably better at Hindi and Hinglish than a US competitor will bother to be. Neither is a permanent moat. Both are eighteen-month head starts, which is what a pre-seed is actually buying.
16:50Twelve months
Four quarters, and a gate on each one.
Every quarter ends with a number that decides whether the next one gets built. The gates are the point of the plan.
Q1 · Prove the pipeline
Capture, transcribe, extract, search
- iOS Share Extension and Android intent
- Transcription + on-screen OCR + structured extraction
- Bulk import from the Instagram data export
- Semantic search over the library
- 1,000 founding members sold
Gate · 60% of imports produce a record the user rates useful
Q2 · Prove the value
Focus mode, exports, paywall
- Focus mode checklists and grocery lists
- Exports to Calendar, Reminders, Notion
- Map view — pins, no nudges yet
- Paywall live, both currencies
- North-star instrumentation shipped
Gate · 2%+ free-to-paid, D30 retention above 22%
Q3 · Prove the habit
The nudge, and Hinglish
- Background geofencing + arrive-and-remind
- Date parsing, calendar sync, two-stage reminders
- Fine-tuned Hinglish transcription
- Android feature parity
Gate · nudged users retain measurably better than un-nudged
Q4 · Prove it compounds
Shared brains and Teams
- Shared folders and merged itineraries
- Public web pages for every checklist
- Teams tier and seat billing
- Series A data pack assembled
Gate · 0.35 installs per paying user per month from shared links
Geofencing sits in Q3 on purpose. It is the most valuable feature and the one that asks for always-on location — a permission users grant to apps they already trust. Asking in week one would cost us the install; asking in month seven, after the library is full, is a different conversation entirely.
17:45Risks
Seven things that could kill this, ranked honestly.
Users won't change the habit — tapping Instagram's bookmark is free, sharing to us is a second step.
This is the real risk, not the API. Mitigation: bulk history import means the app is full and useful before the habit is asked for, so the first value arrives at zero effort. Then only new saves need the share, and we measure share-rate weekly as the leading indicator of everything else.
Churn — subscription utilities bleed, at a 13–14% median monthly rate.
Annual-first pricing, because cheap annual plans retain up to 36% at twelve months against 6.7% for expensive monthlies. Then the structural bet: the nudge creates recurring value where a note-taking app creates none. If Q3's gate fails, this is where it fails.
Platform terms and creator pushback on processing someone else's video.
User-initiated only, transcript-and-thumbnail retention with no video stored, deep links back to the original, creator handle preserved, published opt-out. We keep counsel on retainer from month one — it is a line in the use of funds, not an afterthought.
Instagram ships this itself.
They have had the saves and the compute for years and have shipped a grid. Their incentive is time inside the app; ours is getting you out of the door and to the restaurant. The honest version: if they ship it, we are cross-platform and they are not — and we would rather be the one that also handles TikTok and Shorts.
ReciMe or a funded competitor expands out of recipes into everything.
Likeliest competitive path. Answer: get to places, geofencing and Hinglish before they do, and be the app people already have their library inside. A library is a much better defence than a feature.
Transcription quality on Hinglish and noisy audio makes output feel dumb.
Documented: 9–14% WER on Hindi, past 30% with strong accents, plus 30–50% degradation on code-switched speech. Fine-tuned model, OCR as a second signal, and a hard product rule — when confidence is low we show the transcript and say so rather than inventing a confident summary.
AI costs eat the margin.
A third of a US cent per reel, with two priced-out failover providers and inference costs still falling. Credits and tier caps bound the worst case. This is the risk everybody asks about and the one that matters least.
18:40Who builds it
Two founders, one hire, and no ambiguity about who owns what.
Founder · product & engineering
The pipeline
Share extension, ingestion, transcription, extraction quality, search, the data model. Owns the answer to "is the output good enough", which is the only quality question that matters in year one.
Full-time from month one
Founder · growth & content
The demand
The short-form demo engine, the build-in-public channel, wedge communities, launch days, and the paywall experiments. Owns installs, conversion and the share-link loop. This is a production job as much as a marketing one — eight videos a month, every month.
Full-time from month one
First hire · month four
Mobile engineer
Background geofencing, battery behaviour, push scheduling, offline state, Android parity. The Q3 nudge is the feature the business rests on and it is genuinely hard native work — it should not be a side project for whoever is less busy.
Salaried, not equity-only
Internal — settle before this deck goes to an investor
- Equity split, in writing, with four-year vesting and a one-year cliff for both of us.
- Who has the final call on product when we disagree — named, not "we'll discuss it".
- What each of us is putting in before the raise closes: hours per week, and cash if any.
- What happens if the Q2 gate fails. Agreeing the exit condition while we are optimistic is the only time it can be agreed fairly.
19:20The ask
$300,000 for eighteen months and one honest answer.
Eighteen months is long enough to get through all four gates and short enough that we cannot hide from them. What the money buys is not a launch — it is the retention data that tells you, and us, whether the nudge changes behaviour. That is the only question worth $300,000.
Use of funds
What we come back with
- 250,000 installs, 10,000 paying
- $400K ARR run-rate at 90%+ gross margin
- Monthly churn under 10%, against a 13–14% median
- North star — saves acted on — at 12% or better
- A measured retention lift from geofenced nudges, or a documented reason it didn't work
Terms we're proposing
$300,000 (₹2.5 crore) pre-seed on a standard instrument, priced or convertible, with a board observer seat and monthly written updates including the north-star number whether or not it is good.
If you're not writing a cheque
The three introductions worth more than the money: a consumer-app growth operator who has taken an India-built app to US pricing, a mobile engineer who has shipped background geofencing at scale, and counsel who has argued platform-terms questions for a consumer product. Any one of those changes our odds more than the last $50,000 would.
And the smallest possible ask: give us eight weeks. The 25% intent-saver assumption on slide 14 either survives contact with a thousand real imports or it doesn't, and that single measurement is worth more than the rest of this deck.
20:00Appendix · every number, with its confidence
Every number in this deck, with its source and how far to trust it.
Figures are as of August 2026 and were checked against primary sources where one exists. High means a company filing, an official pricing page or a first-party statement. Medium means a credible secondary source. Low means treat it as a hypothesis — there is exactly one of those, and it is flagged on the slide it appears on.
- 01 Instagram 3.0B MAU · Reels 200B plays/day · Reels $50B+ run-rate — Meta Q3 2025 earnings call.High
- 02 India 480M / US 181M / Brazil 147M Instagram users — Statista country ranking.High
- 03 Reels ≈50% of time on Instagram; 46% of US time in 2025 vs 37% in 2024 — CNBC, Jan 2026.Medium
- 04 177 videos watched per Instagram user per day · 80+ min/day on short-form video — industry aggregators, 2025–26.Medium
- 05 40% of 18–24s use TikTok/Instagram over Google to find lunch — Prabhakar Raghavan, Google, citing internal research, via Nieman Lab.High
- 06 Food, drink and travel lead Instagram save rates · 35M posts across 447,613 pages — Socialinsider benchmarks.High
- 07 88% of interviewees could not recall how to reach Saved posts — independent UX case study, small sample, unreplicated.Low
- 08 Pocket shut down 8 July 2025, archives deleted 8 October 2025 — TechCrunch; Mozilla statement.High
- 09 Instagram API with Instagram Login — four scopes, Business/Creator accounts only, no saved-media scope — Meta developer docs.High
- 10 Private collections endpoints returning 404 / "dead endpoint … unshipped" — public issue tracker, 2026.Medium
- 11 Whisper $0.006/min · gpt-4o-mini-transcribe $0.003/min — OpenAI pricing.High
- 12 Deepgram Nova-3 $0.0043/min batch — Deepgram pricing.High
- 13 Gemini 2.5 Flash-Lite $0.10 in / $0.40 out per 1M tokens — Google AI pricing.High
- 14 Places Essentials $5/1,000 · universal $200 credit retired March 2025 · per-SKU free tiers — Google Maps Platform docs.High
- 15 15% store commission — Apple Small Business Program under $1M, enrolment required; Google automatic on first $1M, subscriptions 15% from day one.High
- 16 2.18% freemium vs 12.11% hard-paywall conversion · 13–14% median monthly churn · ~30% of annual subs cancel in month one · cheap annual retains 36% at 12 months vs 6.7% for expensive monthly — RevenueCat, State of Subscription Apps 2025.High
- 17 CPI $4.70 iOS / $3.70 Android · CAC per paying user $20–80 — user-acquisition benchmark compilations, 2025–26.Medium
- 18 ReciMe $59.99/yr, 5 free imports/week; "10 million users" is the company's own claim — recime.app.Medium
- 19 Wanderlog Pro $39.99/yr · mymind $13/mo · Raindrop Pro ~$3/mo · Readwise Reader $9.99/mo annual — vendor pricing pages, Aug 2026.Medium
- 20 Cal AI — reported $30M+ ARR inside two years, 8.3M downloads by July 2025, acquired by MyFitnessPal December 2025 — TechCrunch, CNBC.Medium
- 21 India ~$0.03 revenue per download vs $0.20+ in SEA/LatAm · record $345M Indian consumer app spend in Q2 2026, +35% YoY · non-gaming 68% of Indian app revenue in H1 2026 vs 58% three years earlier — Sensor Tower via TechCrunch and Business Standard.Medium
- 22 Whisper WER 9–14% on Hindi, ~18% on clear Hindi speech, 30%+ with strong regional accents; code-switched speech degrades WER a further 30–50% — published ASR benchmarks and code-switching literature, 2025–26.Medium
- 23 Unverified and deliberately excluded: any figure for how many saved posts exist per user, or what share are ever reopened. Meta does not publish it and no credible third party has measured it. The 25% intent-saver rate on slide 14 is our assumption, and slide 14 says so.Our assumption