train-mate

Train Mate

An AI-assisted product built with Vercel AI SDK: programme changes validated and reviewed as diffs, never auto-applied. Used by two athletes.

Private repo

2026

next.jstypescriptaitursovercelsqliteauth

Watch the demo ↓

Why

I built Train Mate for two of us training together with different schedules and constraints. We needed one place for programmes, shared workouts, and individual progress, without a spreadsheet or repeated copy-and-paste.

AI can draft programme changes from athlete feedback, but each one stays a proposal until someone reviews and approves it.

How it’s built

I integrated Vercel AI SDK on the server with OpenCode model providers. Feedback becomes a structured JSON proposal: the model can request clarification or draft a patch, Zod validates the result, and malformed output gets one repair attempt. Proposed changes are shown as a diff for explicit approval or rejection. The model has no file, shell, or web tools; application checks protect plan ownership, versions, and safety rules.

Each participant has an independent schedule and prescription, with shared sessions stored once and referenced per profile so history and completion state never collide. Passwords are scrypt-hashed and sessions use opaque HttpOnly cookies. Unsafe cross-origin requests are rejected and repeated failed logins are throttled. Feedback and programme jobs are stored in Turso before idempotent, retryable delivery to Vercel Queues, so no production behaviour depends on an in-memory promise.

AI-drafted programme patches are schema-validated, repaired once if needed, and shown as a diff for approval. They are never auto-applied.
Independent per-athlete schedules sharing canonical workout records, so history never collides between profiles
Scrypt-hashed auth, opaque HttpOnly sessions, throttled logins, cross-origin request rejection
Idempotent, retryable Vercel Queue delivery for AI jobs and programme writes
Mobile-first month calendar with scheduled/in-progress/complete/skipped/overdue states
Per-set logging (reps, load, distance, time, RPE) with automatic PR detection against session history

Demo

Demo: athlete feedback becomes a validated proposal, reviewed as a diff and approved0:46 runtime

Screenshots

Train Mate Today screen with a missed workout needing a decision, the active programme and this week at a glance — Today: a missed session asks for a decision instead of silently shifting the plan

Train Mate programme builder at the Ask stage, with athlete selection and a plain-language brief for a HYROX build — 1 · Ask: describe the goal in plain language; the builder moves through Ask, Clarify, Review and Activate

AI Coach feedback form scoped to the programme, asking to swap a wall-ball finisher for sandbag lunges after shoulder tightness — 2 · Coach request: athlete feedback, scoped to programme, week or day

AI Coach job in progress: the primary model attempt is unauthorised and the job continues on the fallback model — 3 · Durable job: the primary model failed, so the job moved to the fallback model

Review before applying screen showing a validated structured patch, the model attempts and the coach proposal — 4 · Validated patch: schema-checked output, with every model attempt recorded

Workout-level diff showing the added sandbag lunge and removed wall ball for one athlete only — 5 · Diff: exactly what changes, and for whom, before anything is applied

Open questions from the coach with an answer box and options to send answers, reject or revise — 6 · Open questions: the athlete answers, rejects or revises; this proposal was rejected

Train Mate plan calendar for September 2026 with an undo-last-change control — 7 · Plan: the calendar reflects only approved changes, with undo

Links

Live on Vercel for two athletes, verified September 2026. A personal household product with reviewed AI changes and durable workout history; broader adoption is not claimed. Source is private.

Back to the shelfBrowse all projects