← Selected Work

ADC Ready

Voice-AI exam preparation · Design, build & operations

Clinical exam preparation for overseas-trained dentists sitting the Australian Dental Council (ADC) Accreditation Examination — an exam with an 11% pass rate. Candidates hold a spoken consultation with an AI simulated patient in an exam-style station, then receive an ADC-like assessment plus detailed coaching on how to improve.

adcready.com.au →

Realtime

Low-latency spoken roleplay, browser to model over WebRTC

3 AI roles

Patient, examiner and coach, deliberately siloed to ensure single-task focus

45-point

Assessment according to official ADC criteria with full candidate progress

Highlights

Real-time spoken roleplay. Candidates practise realistic patient conversations using natural, low-latency voice interaction.
Three AI roles, deliberately separated. The patient agent stays in character and never tutors; a separate examiner agent grades afterwards from the transcript, the scenario rubric and the professional competency domains. The coach then provides actionable feedback to improve the score.
Examiner-grade feedback, not a chat summary: detailed assessment according to the official criteria published by the ADC, with timestamped quotes from the candidate’s own words and a concrete better alternative for each.
Session recording and playback, so candidates can hear how they actually sounded under time pressure.

Technical detail

React 19TypeScriptViteEleventyNetlify FunctionsDrizzle ORMPostgreSQLOpenAI RealtimeWebRTCBetter AuthAWS S3VitestPlaywright

Realtime voice architecture. Built browser-to-model voice communication using WebRTC and the OpenAI Realtime API, enabling low-latency, bidirectional audio streaming.
Asynchronous grading on background functions with status polling, so a long LLM analysis never sits inside a request.
Self-improving AI workflows. AI-generated scripts are evaluated against defined criteria, refined, and re-tested in a closed feedback loop to continuously improve quality.
Direct browser-to-S3 uploads via presigned URLs — a 15-minute PUT for capture, a 1-hour GET for playback.
Unit and end-to-end tests. Vitest against an in-memory PGLite database that each test seeds itself, plus Playwright end-to-end tests driving the live stack in a browser.

ADC Ready

Highlights

Technical detail

Kaddi

Borderline