Family Wellness

I spent years as a cancer caregiver in my own family. What broke us wasn't the medicine — it was the coordination: three siblings, one group text, a medication management , and a parent who couldn't read any of it. Every caregiving app I tried was built for one user. Care is never one user.

So I built Family Wellness: a multi-party care coordination platform — a senior-facing mobile app and a caregiver dashboard — with an AI assistant, Mena, at the center. And that raised the question that defined this entire project: how do you put an LLM in front of a family's medication data when a single hallucinated dose could hurt someone?

Build Week's answer became my answer: don't trust the model. Verify it.

What it does A senior-first mobile app: a calm home screen that answers "is everything okay?" at a glance, honest medication tracking ("not recorded," never "missed" — the app only claims what it actually knows), appointments, care alerts, and a family care circle. Mena, the AI care assistant, built on a rule most health AI can't offer: GPT-5.6 drafts every response, but a deterministic verifier — plain code, not a model — checks every factual claim against the family's actual database before anything renders. Doses, times, names, schedule states, even whether the record changed mid-verification (version-fenced staleness detection). Anything unverified is blocked and replaced with an honest fallback. The renderer is a pure function over verified claims and human-approved, hash-pinned templates — provider prose is structurally unreachable by the UI. The demo's money shot: we make GPT-5.6 fabricate a dose on camera — and the verifier catches it, blocks it, and logs it. Live. How we built it

Codex built this codebase — under adversarial supervision. The workflow: I worked as architect and reviewer; ( using all different models i could get my hands on and prepared a very clear specific prompt to carry the task) I used Codex agents (GPT-5.6) executed every work order in plan mode; nothing merged without proofs. The rules that made it work:

"Reports aren't proof — retained logs are." Every acceptance ran on disposable Postgres stacks with hash-bound (SHA-256) evidence committed to the repo. Guards watch the guards. When we re-scoped a mis-firing statelessness check, we shipped an adversarial suite proving the guard still catches real violations — 8/8 synthetic attacks (including a memory-write smuggled through the trusted path) blocked. We audited ourselves, adversarially, three times. An independent audit agent ran a 79-case competition proof harness against the tree: first run 0/79 implemented → 28 passing with zero failures → gaps closed in sequence. The FAIL reports are in the repo. We think shipping your own adversarial audits is the point.

The verified pipeline: scope-gated fact bundle → GPT-5.6 generation (structured claims + prose) → strict schema validation → deterministic per-claim verification (exact dose tuples, timezone recomputation, version fences) → completeness gate (every sentence must bind to a verified claim or approved template — unbound prose = blocked) → pure-function render → metadata-only trace ledger (typed columns, no content column — PHI leakage is impossible by schema).

Challenges we ran into "Done" that wasn't. Mid-week we discovered accepted work passing its checks only because of uncommitted files in the dirty tree. We froze, reconciled the whole tree against clean source, and found five latent defects — including a safety guard that had been silently red for days. The fix discipline (never weaken a guard to make it green) became the project's constitution.

I lost my work because of a cyber security warning which I had no idea luckily the work was deployed. Honest vocabulary. The system knows a dose is not recorded — it does not know it was missed. That is the difference. Senior-first accessibility, for real. A 59-route audit, 48 defect classes fixed — max-text layouts, screen-reader contracts, one reduce-motion mechanism, color-independent states — because our primary user is 78 and deserves better than a compliance checkbox. Accomplishments we're proud of

A working two-surface product; a deterministic verifier with ~19 proof classes of retained evidence; a redesigned senior home screen; an AI assistant that is architecturally honest; and the whole thing built solo-founder + Codex in a way we can prove, commit by commit.

What we learned

Agents are extraordinary builders and unreliable narrators — the same is true of LLMs in production. The answer to both is identical: verification as architecture, not vibes. Codex + deterministic acceptance gates built this system; GPT-5.6 + a deterministic verifier runs inside it. It's the same pattern at two scales, and it's how we think consumer health AI should be built.

What's next

A family pilot (recruiting now), provider-compliance completion for real-data operation, and the employer caregiver-benefit channel. The verifier is model-agnostic by design — GPT-5.6 today, swappable by config, guarantees unchanged.

Built With

• ai
• chat-gpt