01Mindfocus
Turning a fragmented learning product into one clear mastery loop
Open Open page
ProductionWeb app
MVPWeb app02PitchArena
Overview
Turning a voice-interface brief into a working AI training product
I took PitchArena from a time-boxed design assignment to a working SaaS MVP, defining the product loop, designing the experience and building the full-stack AI system with a governed agent workflow.
12
playable challenges
59
product-rule tests
40
Judge evaluation cases
Context
A polished voice interaction, but not yet a product
The assignment asked for a domain-neutral voice conversation interface with production-ready polish. Simulated AI was allowed and no real backend was required. The first prototype proved the interaction, but it did not yet define what users should achieve, how success could be trusted or why they should return.
Starting point
The brief ended at the quality of the conversation interface
A realistic exchange had no observable success condition
Feedback and progression were neither persistent nor authoritative
Completing one session created no strong reason to replay
Mechanic
A challenge turns conversation into deliberate practice
Every round places the user in a situation with a resistant persona and observable objectives. The system records the conversation, validates exact evidence, explains what changed and turns the result into a reason to retry or progress.
Choose a Path
Enter a Challenge
Face the Persona
Prove Objectives
Review Evidence
Retry
Decisions
Turn open conversation into an objective-based challenge
The product moved from selecting a scenario to entering a structured challenge. Users now understand the situation, the persona's resistance and the objectives they must demonstrate before they start speaking.
Scenario Selection — Before
Challenge Selection — After
Separate the performance from its evaluation
The live experience now protects the conversation while a separate Judge evaluates user turns against explicit rubrics. This keeps the persona believable and makes progress dependent on what the user actually said.
Context-Heavy Session — Before
Focused Training Round — After
Make feedback the beginning of the next attempt
Results moved from a terminal report to a progression loop. Transcript-backed evidence, missed opportunities, Personal Bests and the next challenge give users a concrete reason to replay and improve.
Performance Report — Before
Progression Loop — After
System
A full-stack AI product whose decisions can be verified
I designed the product and implemented the interface, server routes, data model and AI behavior. Agents accelerated exploration, implementation and review, while product rules, acceptance criteria and release decisions remained under direct human control.
Actor, Judge and Director
The Realtime Actor only performs the persona. A server-side Judge validates objectives against exact transcript evidence. A deterministic Director applies trust, pressure and end-of-round rules.
Server-authoritative rounds
PostgreSQL persists users, rounds, conversation turns and Judge decisions. The browser cannot invent scores or progression, and an unavailable evaluation produces no replacement result.
Governed agent delivery
Work was divided into bounded product, build and review tasks. Code, data, authentication and critical product rules were checked before integration, with human approval at decisions that changed the product promise.
Tests and AI evaluations
Fifty-nine deterministic tests cover product and entitlement rules. A forty-case golden set tests the real Judge across all twelve challenges, including clear success, failure and adversarial behavior.
Production foundation
TanStack Start and React power the interface and server routes. Prisma and PostgreSQL/Neon persist the product state, Clerk handles identity, Stripe supports Free and Pro access, and PostHog measures the complete round journey.
TanStack Start · React · TypeScript · Prisma · PostgreSQL / Neon · OpenAI Realtime · Clerk · Stripe · PostHog
Evidence
A working MVP ready for real user validation
PitchArena now has a complete product and technical loop that can be exercised, inspected and improved. This is delivery evidence, not a claim of product-market validation.
Three Paths and twelve playable challenges with distinct personas
Real-time voice, transcript-backed objectives, feedback, replay and mastery progression
Persistent server-authoritative rounds, authentication, entitlements and analytics
Fifty-nine deterministic product-rule tests and a forty-case Judge evaluation set
Repeat usage, willingness to pay and market demand remain to be validated with real users
03AoMars
Turning focus into a shared effort to restore a world's memory
Open Open page