hi, I’m Mihai!
I’ll keep it nice and quick. Currently I’m a founding engineer/partner at Polarity. I’m extremely obsessed with startups, and don’t see myself anywhere but there or in ventures.
I entered this world a year ago, and I’ve split the last year of my life into startup life and venture life.
- Production traces·Replay every agent decision with full tool calls, prompts, and tokens
- Eval runs·Score outputs across factuality, latency, cost, and custom rubrics
- Sandbox specs·Pin agent behavior with deterministic specs that fail loud on drift
- Regression alerts·Catch model drift and tool failures before they reach your users
- Experiments·A/B compare prompts, models, and policies on real production traffic
- Scorer library·Built-in scorers for factuality, exact-match, file presence, and more
- Slack alerts·Get pinged when an agent regresses or a critical scenario fails
- LLM judges·Auto-grade open-ended outputs with calibrated judge models
- Datasets·Curate golden datasets from production and replay them on every change
- SDKs·Python, TypeScript, and Go SDKs to instrument any agent in minutes
Scanned 14,213 production traces from the last 24 hours across 6 agents, 3 model versions, and 2 sandbox specs.
Regressions detected
- 1. refund-bot v3: Factuality dropped 12% after the GPT model bump — 41 affected traces.
- 2. triage-agent: Tool-call latency p95 spiked 2.4× on the new retriever.
- 3. summarizer-v2: Judge score fell below the 0.85 threshold on legal corpus.
- 4. paragon-hello-world: A new failure mode hit 8% of traces — escalating arguments mid-loop.
- 5. api-endpoint-eval: Exact-match dropped from 91% to 79% on the v3 prompt.
- 6. review-agent-eval: Cost per task up 38% from token bloat in the new system prompt.
What would you like to work on?
Framework Ventures
Investment Analyst at Framework, looking at Seed and Series A/B. I'm also working on the internal agentic tooling the team runs on.
Polarity
Founding engineer/partner at Polarity, acquired by Wander in 2026. Best team I've worked with. I joined at the very start and got to build most of it alongside the best people.
Northside Ventures
Currently I'm an AI Product Engineer at Northside. I joined because of the team, and because of the mix between venture and engineering.
Manulife
Business Intelligence Analyst for a winter term, working on data pipelines and reporting. My first real look at how a large organization moves.
Waterloo & Laurier
School.
Framework Ventures
Investment Analyst at Framework, looking at Seed and Series A/B. I'm also working on the internal agentic tooling the team runs on.