ToCoach

The problem it solves

Sales skill only improves one way: realistic practice with expert feedback. But human role-play coaching costs $75–200 an hour, has to be scheduled days out, and the written debrief lands 30+ minutes after the moment that mattered. The solo founders who most need reps get the least access to them.

And there's the part nobody solves: accountability. When practice is uncomfortable, people avoid it — they browse the dashboard, reread lessons, and quietly stop doing sessions. No coaching product notices. This one does, and it responds.

What I built

A local-first desktop app where you talk — out loud — to an AI prospect that pushes back like a real buyer: it interrupts, drops competitor names, and escalates when your answers are weak. Speech recognition and synthesis run on-device; the only external call is the LLM, and even that routes through a subscription CLI so marginal cost is zero. After every session, a fleet of LLM pipelines runs the entire coaching back-office — scoring, debriefs, curriculum, buyer research — while a coach brain maintains a deep model of the student and personalizes everything that follows.

Key features

• Role-play that feels adversarial — 16 scenarios × 12 personas × 3 difficulty tiers (500+ configurations) with adaptive pressure tactics that escalate on weak answers and ease on strong ones.
• Feedback in seconds, not next week — rubric-scored drills, generated exams (0–100), and 4-dimension debriefs with line-by-line coach rewrites, delivered the moment a session ends.
• A coach that remembers you — a ~44-field deep student profile (confidence baseline, fear patterns, price confidence, objection habits, core internal blocks) updated after every session and injected into every future prompt.
• Accountability with teeth — five escalating response levels driven by real presence data, per-feature privilege locks, and behavioral pattern detection that catches avoidance before it becomes quitting.
• Self-improving curriculum — every session generates vocabulary and communication insights; every 5th session authors a new scenario variant targeting current weaknesses; every 14th session the coach audits its own strategy.

System internals

Impact

Reliability & safety engineering

• The never-brick rule: no blocking UI ever waits on a model call. Every LLM call runs in an isolated child process with a 90-second hard timeout, and heavy post-session generation is deferred off the interaction path entirely.
• 5-stage JSON recovery: structured outputs survive messy model responses — strip code fences → direct parse → greedy outermost-braces match → all candidate blocks largest-first → fail loudly with a truncated preview. Imperfect output degrades to a retry, never a broken screen.
• Deterministic fallbacks with authority: every accountability overlay, punishment, and reward renders from deterministic templates — the discipline system keeps working when the AI is down, which is exactly when a user is most likely to test it.
• Injection-safe by construction: prompts travel over stdin (never argv or a shell), the deep-profile writer accepts only whitelisted columns, and the renderer is sandboxed behind the contextBridge with no direct Node access.
• Privacy as architecture: transcripts, scores, and the student profile live in a local SQLite file on the user's machine. Voice never leaves the device; the only network egress is the LLM prompt itself.