Yash Savalia
I'm a founding engineer at Bluejay in San Francisco, where I work on evals and research for voice AI.
Lately that means benchmarks: MIVAS for speech-to-speech models, MISTT for speech-to-text, and CALLM for the LLMs behind phone agents.
Before Bluejay I studied physics, computer science and astronomy at Illinois.
Outside of work I'm learning Japanese and building small tools for myself.
§1 Work
Table 1. Things I have built.
- Bluejay
Testing and observability for voice and chat agents. Simulated callers before launch, evals on live calls after.
2025 → now - MIVAS Bench
Benchmark for speech-to-speech models on multi-agent phone calls in healthcare, legal and customer support.
2026 - MISTT Bench
Benchmark for speech-to-text models: accuracy, latency, price, and how they hold up on bad phone lines.
2026 - CALLM Bench
Benchmark for the LLMs behind phone agents, run over a simulated phone line.
2026 - kotoba
Video player with a hover dictionary and Anki card mining, for learning Japanese from your own media.
2026 - CloudPCap
Packet capture for measuring the latency of trading bots on AWS.
2024
§2 Log
- Bluejay Founding AI engineer, San Francisco
Voice AI evals and research.
- Disruption Lab at Gies Technical lead
Risk and profitability models for an auto lender.
- University of Illinois Urbana-Champaign
B.S. Physics; Computer Science + Astronomy.
§3 Principles
- Find the cause, not the symptom.
- Read it before you change it.
- The best code is the code that never gets written.
- Build it, then argue about it.
- Short sentences.