This is my submission to the Hacktoberfest Open-Source AI Challenge, Week 3. The theme is Slow and Steady.
You answer one question a week. Ninety seconds, by voice. purpoise does the rest: it reads the signals you already generate, and tells you honestly whether your work carries your purpose, partly carries it, crowds it out, or has nothing to do with it.
Ikigai is not your job. Most apps about purpose get that wrong, and then quietly turn into career coaching in a kimono.
What I built
purpoise keeps two models. One is your purpose. One is your work. A verdict compares them. The purpose model is built with no reference to an employer, a job title, or an industry. Work is inferred from work-channel signals, and only those. Mixing the two is how a job gets mislabelled as a reason for being.
The week looks like this. You open the app. It asks one question, chosen because a circle is still thin. You answer. The rest of the week it reads what you already produced: a purchases export, bookmarks, a social download, mail, a calendar, commits. It redacts, extracts signals, and files each signal under love, good-at, needs, or paid. A trait that only ever appears at work is hatched and capped until it shows up in a second channel. The verdict says which circles are present, which are missing, and which claim sits near a miss. Every claim opens onto dated evidence. Reject one and the verdict is computed again.
The person who gets to be lazier is the one who would otherwise take a fifty-question purpose quiz and still not know whether Tuesday's work had anything to do with the answer. The quiz is replaced by one question and a pile of records they already have.
The client is live at https://jedau.github.io/purpoise/. That Pages build is the last push to main. The walkthrough below was recorded from the local client, which is ahead of that push.
The file is docs/demo.mp4. It is 26 seconds, silent, 1280×800. The blob link works once this file is on main. It is not embedded in this draft, because the draft was written before that push. It is also not the three-minute shot list in docs/DEMO.md. That list still includes airplane mode, a live redaction, and a twelve-week replay. None of those are in this recording.
What the recording shows, in order:
- The four circles overlap. The middle reads Ikigai, and a purpose sentence marked as a reading.
- A tap on love opens sketching, night walks, and Sunday soup. A second tap puts the middle back.
- A tap on good-at opens letters, and a hatched Tuesday status report marked work only.
- The ledger. Teaching is first, with
[teach-1]and[teach-2]. Thursday briefings is the paid circle, with[brief-1]and[brief-2]. - The verdict opens on "Your work carries it." The middle is Ikigai. Passion, Mission, Profession, and Vocation are all present.
- The page moves down the four questions. Love holds sketching, night walks, and Sunday soup. Good-at holds letters, and a hatched status report marked work only. Needs holds teaching. Paid holds the Thursday briefings.
- Reject on teaching. The card says the verdict will move.
- The verdict again: "Your work carries part of it." The middle is open. Mission and Vocation are gone. Needs says it is missing. Paid still holds the Thursday briefings.
- Consent, as three facts and three groups: bring a file, sign in, request a file.
- The circles again. The needs circle is empty, and the middle says it is open.
How it is built
Nine stages, in order: source, redact, extract, corroborate, volition, circle map, diagnose, verdict, synthesise. The weekly answer writes labels back into the same file.
Volition is placement. It decides where a signal lands. It does not decide whether the signal counts. The score is a local TabPFN sidecar when that process is up, and a documented heuristic when it is not. The heuristic is 0.5 + 0.2 outside work hours, + 0.15 on a weekend, + 0.15 if you started the thread, clamped, and it cannot fall below 0.5. Work hours are 09:00–17:00 UTC. A weekday work-channel vector sent to the sidecar, with the licence key set, came back near 0.29. Stopping the sidecar flags the span and keeps the heuristic. Both paths are tested.
Text inference is local Ollama. The extract bake-off scored four installed models on four labelled sentences. All four matched every gold kind, so F1 was 1.000 for each. The registry then picked the lower median latency, llama3.2:3b at 926 ms. qwen3:4b was 1333 ms, gemma3:4b 1633 ms, mistral:7b 2086 ms. Cost per thousand was 0, because the process is on the machine. n is 4, so the table is indicative. The 500-record replay was not run. A later call to Backboard authenticated and then refused chat, because that free credit is reserved for Memory and RAG. The refusal was not stored. The table stayed the local run.
The held-out check is a different, smaller thing. Eight rows train a per-subject mode. Two rows are held out. Accuracy on those two is 1.000 against a naive baseline of 0.500. The confidence interval is 0.342 to 1.000. n is 2, so that row is indicative too. The Tinker API was not called. The delta in docs/EVAL.md is that local subject mode against a global mode: accuracy 0.500 to 1.000, macro-F1 0.333 to 1.000, cost 0.
The web client has no on-device model. Mobile is the tier that can redact and extract on the phone, through Flutter Edge AI, and the settings screen says which tier ran. The live agreement against twenty fixtures was not measured. No phone was connected, and the int4 file was not downloaded. The scorer itself is tested: 18 of 20 matching kinds passes, 17 does not.
The API is Fastify. The worker process stays up and does not start Temporal, so the activities run inside the request. MongoDB and Tiger Data are in the design and were not written during this build. The character file the demo shows is the in-app ledger.
What stays on the machine
| Capability | Where it ran | Leaves the machine? |
|---|---|---|
| Extracting a signal from a sentence | Local Ollama | No |
| Choosing the extract model | Local bake-off, four models | No |
| Placing volition | Local sidecar, or the heuristic | No |
| The quarterly letter | Composed in the app. A model reply is kept only inside the word range, and only if it cites at least five known ids | No |
| Ritual voice | ElevenLabs if the key is set, otherwise a local tone and whisper.cpp | The voice does, when that call is made. The Flutter screen still types the answer |
| Web redaction | The server | Yes. The phone tier is the one that redacts before anything is sent |
One local extract run on this hardware cost $0.00. I did not meter the same four sentences against a hosted chat API, so I will not invent a monthly bill. The number I have is the one in the table: $0, because no token was purchased.
Why open weights matter here
Purpose is the most personal corpus a person owns. A twelve-month model of you cannot live on someone else's server, behind someone else's model, on someone else's subscription. You also have to inspect why the app concluded something about your life. Open weights make both possible. Every text step in this build runs on those weights. There is no hosted LLM in that path. ElevenLabs is the one hosted model, and it covers voice only. Backboard is a comparison gateway, not the production extractor.
The four circles need a credit, because they are not traditional ikigai. Andrés Zuzunaga drew a purpose diagram in 2011. Marc Winn relabelled it ikigai in 2014, and Dan Buettner's reporting from Okinawa carried that picture further. Traditional ikigai is smaller and more daily. purpoise keeps the diagram as one mode and a weekly practice as the other. Money is not required. The paid circle may stay empty, and the app is allowed to say the picture is unclear.
The limits, from the privacy note, without the marketing gloss:
- Web has no on-device redaction. Redacted text reaches the server. Mobile does not.
- The stores and the API host were designed as managed accounts. In this build the stores were not written, and the API host was not created.
- A tunnel, when one is used, transits Cloudflare. They see redacted text in flight.
- Production inference, when the API is up, depends on one laptop being awake.
- A search API would learn what you are curious about, at query-string granularity. That call was not made.
- ElevenLabs hears your voice for a ritual answer, when the route is called.
- Backboard saw an authenticated request and refused the chat. It did not receive a production corpus.
- Redaction reduces exposure. It does not eliminate it. Context can still imply a third party.
What I can count
I did not run the app on my own week of mail and spending, so I will not report hours saved or decisions made. The measured use is the fixture set.
The bake-off took four sentences and four local models. Selection was latency, not quality, because quality was tied. The held-out pair is two rows. The sidecar smoke test was one feature vector. On the local client the four circles meet, and the verdict is that the work carries it. The verdict page reads each question and the pairs that both have a trace. Rejecting the teaching claim opens the middle again, drops Mission and Vocation, and the verdict becomes partial. The recording above is that sequence. The engine fixture the tests lock is still the older partial case, with pay missing. That recompute is the lazy payoff I can stand behind: you do not re-score your life by hand. You reject one receipt and the conclusion moves.
Time to payoff on the public client is the time to open https://jedau.github.io/purpoise/ and read the verdict. There is no account required for that demo ledger.
Sessions
A DevRelay session was not exported. An Entire session link was not captured. The decision log is docs/DECISIONS/. The unknowns log is docs/UNKNOWNS.md.
Prize categories
No partner line from the original list passed its checklist, so none is claimed. The entry is the overall challenge. Removed, and why:
- Gemma. The on-device file was not downloaded, and the parity run was not measured.
- TabPFN. The sidecar answers a numeric vector locally. Accuracy against a naive baseline was not reported for that task.
- Tinker. The export and a local delta ran. The fine-tune API was not called.
- Render. The blueprint is in the repo. Creating a service asked for a card even on the free web plan. The API was left undeployed.
- Auth0. Sign-in exists. Token Vault was not called, and there is no step-up gate.
- Backboard. The key authenticates. Chat was refused. The stored scores are the local bake-off. The assistant does not use it for memory.
- ElevenLabs. The server route can speak and transcribe. The screen in the client still types.
- Browserbase. A download the person already has is read from a file. No browser session is opened.
- Entire, GitHub Copilot, Mastra, MongoDB Atlas, Sentry, SerpApi, Temporal, Tiger Data. The code names the role. The checklist for each one asks for a run this build did not do.
Code
The repo is https://github.com/jedau/purpoise. The licence is MIT.
Locally: copy .env.example to .env, then follow docs/RUNBOOK.md. The web build publishes from .github/workflows/deploy-pages.yml on a push to main. The base path is /purpoise/.
Commits after 2026-10-25 23:59 PDT, if any, are listed in the README.
Top comments (0)