Mettle

A private GPT-5.6 coach that turns real-world social reps and conversations into patterns, memory, and the next move to practise.

Comment

Inspiration

Dating apps can produce matches or silence, but they rarely help someone understand what happened or build confidence for the next real-world interaction. Mettle turns social confidence into a private practice loop: do a rep, reflect, find the pattern, choose the next move.

What it does

Mettle is a private Android social-confidence coach for men. A user can log an interaction or, with explicit consent, record a conversation. GPT-5.6 turns that event into a structured rep, updates people-aware coach memory, surfaces useful patterns, and recommends one concrete action to practise next.

The product includes:

  • Guided rep capture and reflection
  • A real coach chat with scoped, editable memory
  • People profiles that preserve relevant context
  • Conversation recording, playback, recovery, sharing, and opt-in AI analysis
  • Chat- and audio-to-rep extraction
  • Goals, patterns, and next-action recommendations
  • Durable on-device storage, backup/restore, export, and delete controls
  • Encrypted on-device storage for the user's OpenAI API key

How we built it

Mettle combines a compact HTML/CSS/JavaScript application with a Capacitor 8 Android shell and a native Java recording service. The app uses the OpenAI Responses API with GPT-5.6 for coaching and structured memory/rep extraction, plus OpenAI audio transcription for consented recordings.

Codex was the engineering partner throughout Build Week. We used it to turn product intent into OpenSpec work, implement the web and native Android paths, review failure modes, build deterministic test harnesses, and repeatedly verify clean Android builds.

Why GPT-5.6 matters

A generic chatbot is not enough for this problem. Mettle needs to reason across a user's latest interaction, their chosen person, prior memories, goals, and privacy settings, then produce a grounded next step without overwriting newer user edits. GPT-5.6 powers the coach and the structured extraction that makes this loop useful rather than merely conversational.

Challenges

The hardest work was making recording reliable across Android lifecycle events, preventing stale AI responses from overwriting newer state, keeping memories scoped to the right person, and ensuring consent and privacy were explicit at every upload boundary.

Accomplishments

  • A coherent log → reflect → pattern → next-rep product loop
  • Native background audio capture with recovery and a real recording library
  • People-aware coach memory with edit and delete controls
  • Guarded GPT-5.6 write-back using epochs and tombstones
  • Durable local state, migrations, backup, and restore
  • Runnable Android packaging with targeted verification harnesses

What we learned

AI coaching becomes trustworthy when privacy, memory, and action boundaries are product features—not afterthoughts. We also learned that Codex is most effective when every substantial change has a narrow, runnable check and an adversarial review pass.

What's next

Next we will distribute a signed Android beta, add an optional on-device transcription path, and run structured user testing on whether Mettle's recommended next actions improve real-world follow-through.

Build Week extension

Mettle existed as an early prototype before the event. During the submission period it was meaningfully extended with persistent local state and migrations, the remembered coach, people profiles, Capacitor Android packaging, durable native storage, background recording and recovery, GPT-5.6 Responses API support, transcription, structured memory extraction, and chat/audio-to-rep capture. The dated repository history and Codex sessions document this work.

Built With

Share this project:

Updates