Founding Software Engineer · Repovive YC F'26
Repovive is a Y Combinator-backed science expert community where experts solve unsolved scientific problems, starting with theoretical computer science and math proof reasoning. It hosts live programming and math contests and has grown to 10,000+ users.
- Joined as the first engineer and built the core systems from the early days through YC.
- Onboarded and supported new engineers, senior hires included, guiding them through the codebase and projects; mentored junior programmers to improve their first code deliveries.
- Moved CI to Vercel when the GitHub Actions quota ran out: one fail-fast pipeline (format → typecheck → test → build) with no production secrets.
- Tracked down live-contest outages using Claude Code, MCP servers, PostHog and log analysis, then shipped the fixes.
Repovive Judge & Contest Infrastructure
Led design and build of the distributed judge (judge1, on Judge0): multi-language compilation, isolated sandboxes, and agentic test generation.
- Scaled submissions 10× (50 → 500+ concurrent) with a Redis-backed async pipeline.
- Cut infra cost 95% ($600 → $25/mo) by moving off serverless to Docker / Nginx / Redis.
- Sandboxed 20+ languages in resource-limited Docker images.
- Redesigned the DB layer for 50 MB+ contest test cases.
Repovive AI Interviewer Platform
Co-built a real-time voice AI interviewer for coding, recruiter and behavioral interviews, used for both practice and B2B screening.
- Built "Kevin", the AI coding interviewer: 6 LLM tools, live Monaco editor reads, and judge-backed code runs.
- Made interviewers markdown-defined and versioned, so admins add new ones without code changes.
- Launched B2B screening: single-use invite links, auto-grading, and a partner ranking dashboard.
- Hardened WebRTC voice turn-taking (1s to first word) across 7 LLM providers.
- Shipped credit-based monetization with free attempts and auto-refunds.
Repovive Build v2: AI Problem-Authoring IDE
Co-founded a real-time collaborative IDE for competitive-programming problem setting, backed by autonomous agent pipelines.
- Built an event-driven agent workflow with 28 tools and streaming turns.
- Secured agent actions with execution locks and ~1s atomic cancellation.
- Kept 10-minute streaming agent sessions stable.
- Scaled concurrent server capacity 10× for automated test runs.