🧬 Harness Health Engineering

A local AI lab that helps you discover what actually improves your life.

Physiology streams in from Whoop · lived experience goes in by text · voice · selfie · an on-device agent runs n-of-1 experiments and tells you what actually makes life better — all as plain Markdown you own.

↑ a black-and-white product page · cryptoyoginya.github.io/harness-health-engineering

🧭 Explore

The one idea

Every health gadget you've owned optimised a number. Recovery. HRV. Steps. And somewhere along the way the number became the point, and your actual life — whether you felt good, did meaningful work, saw people you love — fell out of frame. That's Goodhart's law wearing a fitness band: when a measure becomes the target, it stops measuring anything that matters.

This project flips it. The top-level metric here is not a body score. It is one honest question:

⭐ North-star: "Is my life actually better?"

Everything else — recovery, sleep, HRV, supplements, training load — is demoted to what it really is: an instrument in service of that question. The system will happily tell you that your body looks great this week and your life doesn't, and then help you fix the right thing. No wearable can say that, because no wearable knows what your good life looks like.

Built for people who have tried enough supplements, trackers, routines, and protocols to know that the hard question is not “what is optimal?” but “what is optimal for me, in the life I actually live?”. Not for people who want motivation, streaks, badges, or a prettier sleep chart.

Related MCP server: MCP Index Notes

The engine: n-of-1 experiments 🔬

Correlations are guesses. "You sleep worse when you drink" can't tell you if the drink did it or if a hard day caused both. So instead of guessing, the harness runs n-of-1 trials — single-subject experiments, the real methodology personalised medicine uses to decide if something works for one specific person. Every experiment is pre-registered:

This is the difference between "I tried a thing once" and knowledge that compounds. The payoff is a sentence no tracker will ever give you:

"Creatine moved nothing for you in three weeks — stop paying for it." "Caffeine before 14:00 bought you +35 min of deep sleep — it's a rule now."

One variable at a time. A clock. A criterion. A verdict that becomes a rule. That's the whole game, and it's why this gets smarter every week instead of just logging more.

stateDiagram-v2
    direction LR
    [*] --> proposed: /exp new · or agent proposes
    proposed --> active: baseline + criterion set
    active --> active: /exp extend
    active --> merged: criterion met → becomes a rule
    active --> reverted: missed / stopped → logged why
    merged --> [*]
    reverted --> [*]

You drive it from anywhere: /exp new to start, /exp extend to give it more time, /exp stop to call it — or just ask the agent to design a tighter one. Either way it ends in a verdict, not a vibe.

Why it's different from Whoop's journal

Whoop's journaling is genuinely good — and it has a ceiling. Here's where this goes that it structurally can't:

The discipline is the product. The data is just raw material. The full scientific rationale — n-of-1 design, surrogate-endpoint failure, evidence labelling, confounding — plus a clinician-facing brief with references is in METHODOLOGY.md.

How it works

flowchart TD
    W["Whoop API v2"] -->|"OAuth2 · auto 9:00 · serialized refresh"| SYNC["sync.mjs"]
    H(["You · text · voice · selfie"]) -->|"events · mood · energy · social · work · body"| BOT["Telegram bot<br/>capture layer · on-device"]
    SYNC -->|"recovery · HRV · sleep · strain"| RAW["01_raw · daily record"]
    BOT --> RAW
    RAW --> SRC["02_sources · weekly notes<br/>FACT / INFERENCE"]
    SRC --> SYN["04_synthesis · patterns + life-quality"]
    SYN --> EXP["05_decisions · n-of-1 experiments<br/>one variable · criterion · verdict"]
    EXP --> RULE["CLAUDE.md · rules that stuck"]
    EXP -.->|"merge / revert"| RAW
    NS["⭐ north-star:<br/>is life better?"] --- SYN

    AGENT["on-device agent<br/>RAG · cited synthesis · MCP"] --- RAW
    AGENT --- SYN
    AGENT --- EXP

The loop: Signal → Ingest → Source note → Synthesis → Experiment → Verdict → Rule → repeat. Objective body data arrives on its own; you add a 40-second diary; on Sundays the agent scores your week against the north-star and finds what's actually moving it.

What it tracks

Daily (auto): recovery, HRV, resting HR, sleep, strain — from Whoop. Daily (you, ~40s — by text, voice, or selfie): events (your impressions journal), mood, energy, social, work, movement, supplements. Weekly: one integral "is life better?" 1–5. Quarterly: six life dimensions — emotion, connection, body, meaning, autonomy, growth.

Three ways in 📥

Lived experience is messy, and you shouldn't have to sit at a keyboard to capture it. The bot takes three modalities — mix them freely, several per day:

A photo is auto-classified on-device (CLIP, zero-shot — the image never leaves the machine) into one of three local diaries, each read against its own canon — all descriptive, never diagnostic. No caption needed; a caption simply overrides the guess:

Every message lands as a timestamped line - 14:30 … in 01_raw/health/YYYY-MM-DD.md, so the shape of the day is preserved — not flattened into one average.

A gentle daily rhythm. A morning body brief, a nudge or two with a question rotating across life domains, an evening recap — /quiet mutes the lot. Commands set the rest: /north your direction · /habit what's already standing (15k steps, omega for years — so it's treated as baseline, not re-discovered) · /exp to run, extend, or /exp slip to log a lapse on an n-of-1 (one active at a time — truly one variable, so the effect is attributable) · /week · /month · /year for horizons.

Numbers stay where they belong: only Whoop signals are summarised numerically. Mood, energy, and meaning are never averaged into a score — a single honest "is life better?" is more valid than an arithmetic of parts.

Three photo diaries, one on-device eye 🤳

Your face, your plate, and your gut all leave visible traces of how you live — and a wearable sees none of them. The harness turns photos into longitudinal signals, while staying strictly on the safe side of the line: it describes and compares over time, it never diagnoses.

The on-device eye: how routing works 👁️

Send any photo — no caption, no menu. The bot identifies what it is with CLIP zero-shot image classification running locally (ONNX, the same on-device stack as the voice transcriber). The image never leaves the machine; nothing is uploaded.

flowchart TD
    P["📷 photo<br/>(Telegram)"] --> DL["bot: download"]
    DL --> CLIP{"on-device CLIP<br/>zero-shot · ONNX"}
    CLIP -->|face| S1["photos/ · face"]
    CLIP -->|meal| S2["photos/food/ · food"]
    CLIP -->|toilet| S3["photos/stool/ · gut"]
    S1 --> STORE["local · gitignored<br/>only a filename ref in the daily log"]
    S2 --> STORE
    S3 --> STORE
    STORE -.->|"on request: 'review my …'"| AGENT["agent vision<br/>(strong model)"]
    AGENT --> OBS["dated observations · local"]
    OBS --> HYP["hypothesis"] --> EXP["n-of-1 experiment"]

• Selfie is the safe default — the classifier must clear a confidence margin to file a photo as food or stool; otherwise it stays in the neutral face diary.

• A caption always wins — write "food" / "stool" / "selfie" to override the guess.

• Storage is the only automatic step. Content is never auto-analysed — the deep read happens only when you ask the agent ("review my selfies / food / stool"), keeping a strong model's quality without sending anything anywhere by default. (Warm classification ≈ 150 ms.)

1 · Face — the visual diary

What the agent reads from a face — over time, not in one shot:

2 · Food — read against the plate canon

Method: the plate canon — ~½ vegetables · ~¼ protein · ~¼ complex carbs · + healthy fat.

No calorie counting — unreliable from a photo; the agent reads composition and timing, not numbers. Described neutrally, never moralised — an eating-disorder guardrail keeps the focus on food → how you feel, not control or guilt.

3 · Gut — the Bristol Stool Scale

Method: the clinical Bristol Stool Scale (type 1–7, with 3–4 as the healthy middle) plus colour.

Black/tarry stool or visible blood → see a doctor, not a diary entry — no interpretation attempted.

Hard limits, across all three diaries

• Not a diagnosis — any worrying sign resolves to "see a doctor," never an interpretation.

• Constitution ≠ trend — innate features (dark circles, face shape) are constants, not changes.

• Shooting noise — light, angle, makeup, time of day distort more than physiology; mismatched shots → low confidence.

• Biology lags — skin breakouts (and gut shifts) surface days after a trigger; never pinned to "yesterday."

• One shot ≠ a pattern — value is the trend; strongest evidence = paired shots, "morning after X vs morning without X" — already almost an n-of-1.

Architecture: the knowledge pyramid 🧱

Everything is plain Markdown in a layered pyramid — raw signal at the base, decisions at the top. Each layer only consumes the one below it, so evidence flows upward and nothing is asserted without a trail back to its source.

flowchart TB
    CAP["✍️ 🎙️ 🤳 capture · ⌚ Whoop"] --> R["📥 01 · raw"]
    R --> S["🔖 02 · sources<br/>FACT / INFERENCE"]
    S --> Y["🧩 04 · synthesis<br/>patterns · life-quality"]
    Y --> D["🔬 05 · decisions<br/>n-of-1 experiments"]
    D --> RU["✅ rules that stuck"]
    D -. "merge / revert" .-> R
    NS["⭐ 00 · north-star"] -. "is life better?" .-> Y

The retrieval, hooks, and agent all read this pyramid — never raw guesses. The full scientific rationale lives in METHODOLOGY.md; the layer contracts in AGENTS.md.

Technology

Quickstart

📖 Standing up your own copy? Follow the full step-by-step in SETUP.md — Telegram bot, Whoop app, .env, auto-start, and an honest privacy breakdown.

corepack enable
pnpm run setup
cp .env.example .env            # WHOOP + Telegram bot tokens
pnpm whoop:auth                 # one-time OAuth → .whoop/ (gitignored)
pnpm whoop:sync                 # pull physiology + push a morning brief
pnpm kb:index
pnpm kb:think "what actually moves my life quality?"
pnpm kb:doctor

Daily rhythm: morning sync + brief run themselves · evening text/voice/photo your diary to the bot · Sunday ask the agent to review the week.

Privacy & safety

Local-first, but honest about the edges. Your record lives on your disk; pre-processing (voice→text, photo-type, search) runs on-device; photos and sensitive diaries are git-ignored and never pushed; secrets (.env, .whoop/) are git-ignored and deny-read by the agent. What does leave the machine, by design: messages transit Telegram (the Bot API is not end-to-end encrypted), Claude/Anthropic processes what you send the agent to analyse, and Whoop returns your physiology. Nothing is sold or posted publicly. Full breakdown in SETUP.md.

Not a medical device. The agent never diagnoses; any worrying signal resolves to one recommendation — see a specialist.

Contributing

Built as one person's body-as-codebase, designed to generalise to any self-quantifier. PRs especially welcome on: wearable adapters (Oura, Garmin, Apple Health), a live Whoop MCP server, richer experiment designs, and synthesis evals. See CONTRIBUTING.md and the open issues. (Yes, Bryan, you too.)

License

MIT © 2026 Christina Vinter. See LICENSE.