Slate
Say what you did. It files itself.
Hold one button and talk. Slate transcribes the ramble, splits it into separate activities, works out which day each belongs to, closes todos you already wrote down and pulls out what you spent — all in the background.
Creator — product, AI pipeline & mobile web
- Web Speech API
2026
Private build — demo on request
Logging your day is tedious, so nobody does it. Slate turns it into one gesture: hold, talk for as long as you like, let go.
Transcription runs on Whisper large-v3-turbo and reasoning on an open 120B model, both on Groq’s free tier, with a rotating key pool and a live usage monitor. You are only asked a question when getting it wrong would corrupt the log.
It ships today as a web app that looks and behaves like the Android app, so it works on a phone over the local network, with a native Expo app in progress.
Logging what you did is useful and tedious, so nobody keeps it up. Every tracker wants a form, a category and a date for each entry.
The bet: if logging were one gesture — hold, talk, let go — people would do it. That means turning an unstructured ramble into clean, correctly dated entries, cheaply enough to run on free model tiers.
- Web Speech API
- 01
Capture
Hold to talk. The phone uses Web Speech or sends audio to Whisper; the dev server issues its own certificate so the mic works over the LAN.
- 02
Guard
Whisper invents text on silence — repetition loops or an echoed prompt — so decoder statistics and a prompt-echo check drop it before it becomes 'activity'.
- 03
Cache first
Clauses are normalised (fillers stripped, tense kept) and matched against learned patterns with {n} slots; fully known captures cost 0 tokens.
- 04
One structured call
Only unknown clauses go to gpt-oss-120b in a single call with low reasoning effort.
- 05
Validate and file
Dedup, date resolution, todo closing and the confidence gate run in code; a question is asked only when a wrong guess would corrupt the log.
One-button capture
Hold to talk, release to file — no forms, no categories to pick.
Ramble splitting
One recording becomes separate activities, each placed on the right day.
Todo closing
Matches what you did against todos you already wrote down and ticks them off.
Spend extraction
Pulls amounts you mention into a spending log automatically.
- 01Self-signed HTTPS dev server so the phone microphone works over LAN
- 02Groq key-pool rotation with live usage tracking
- 03Research notes on Android capture surfaces and the path to native
- 04Monetisation and strategy documented alongside the code
One call beats an agent loop
On the same 95-word capture, a 14-call tool loop spent 49,054 tokens in ~17 s and missed an activity; one structured call covered all 12 in ~1 s for ~1,000 tokens. Bounded extraction has nothing to explore.
The cache is the moat
Every resolved clause is learned, keyed by a signature with numbers as slots — so repeats are free and the cache becomes personal, which competitors can't copy.
Conservative caching
A wrong cached entry is silent and repeats, so todos, dated clauses, low-confidence items and clauses that lean on a neighbour are never cached.
Nothing logged that wasn't said
Logging a Whisper hallucination is worse than logging nothing, so transcripts are gated on the decoder's own statistics before anything is filed.
- →Test Whisper properly on the user's own accent and vocabulary — the one risk that gates the native build.
- →Ship as a PWA and measure whether people keep talking to it after week three.
- →Capture surfaces on Android: the notification bar is shipped; a Quick Settings tile and home-screen widget are written and waiting on a build.
- →On-device whisper.cpp for privacy.
Want something like this?
I can build one for you.
Tell me what you have in mind — I'll reply with questions, a plan and a quote.