Aditya Kulkarni

TalkLab

Oral exams run by an AI examiner, so every student in a large course can be assessed in conversation. I built the platform and took it from prototype to its first pilot.

Role
Senior AI researcher
Since
2026
Pilot
1,500 students
Where
U of T Scarborough

Pilot

A real-time voice examiner

I built the voice examiner students talk to, and the tools instructors use to enrol and invite a whole course. It runs on LiveKit, Deepgram, ElevenLabs, Azure OpenAI and AWS.

How a TalkLab oral exam flowsA student talks with an AI examiner in real time. The conversation becomes a transcript, which a scorer grades against the course rubric. The examiner and the scorer share one contract describing what was asked, and that contract is tested on every change.StudentspeaksAI examinervoice, real timeTranscriptwhat was asked and saidScorergrades against the rubricShared contracttested on every changeStudentspeaksAI examinervoice, real timeTranscriptwhat was asked and saidScorergrades against the rubricContractshared
The examiner and the scorer share one contract, tested on every change.

Reliability

Grading that holds steady

The examiner and the scorer share a contract that’s tested on every change. A prompt update can’t change what a grade measures without the tests catching it.

Research

How people behave with AI

We’re researching how people behave with AI systems, and how to reliably assess those interactions across every kind of interface: voice, video and text.

NextBorderPass