Harsh Kasat

Design and run the sandbox layer that executes AI coding agents for thousands of user projects, on Convex + Bun/TypeScript. Multi-provider sandbox architecture: one SandboxProvider interface over Daytona and E2B, with golden templates, a warm pool for instant project start, and automatic failover (e.g. new projects route to E2B when Daytona's org disk cap fills). Lifecycle sweeps that scale: an idle-park sweeper that reconciles against the provider's own running list instead of our rows, parks newest-first, and skips sandboxes a client just woke — fixing a starvation bug that had silently parked nothing for weeks. Fleet-wide job runner: a queue-claim service with jittered claims so hundreds of runners stop racing for the same run, and timeout sweeps that finalize one stuck message per transaction so a bad row can't block the batch. Memory-safe static deploys: moved archive packing from a 512 MB Convex action into the sandbox itself (find | tar | gzip), added size guards and a settings-based rollback switch. Security by default: made Convex functions internal unless explicitly exposed, moved the admin-key hand-out behind a session-gated route, and added a CI gate that fails the build when the public surface grows.

Early engineer on an AI app builder: owned backend services, hosting infrastructure and the agent runtime. Screenshot service built for throughput: BullMQ queue in front of a Puppeteer browser pool with a mutex per browser, per-project+URL throttling, stalled-job detection, 3-attempt retry with 24h re-queue, and idle auto-shutdown so the fleet costs nothing when quiet. Hosting migration: moved user-project static hosting from Freestyle to Vercel with project reuse on redeploy, and cut Convex usage along the way. Usage metering: moved Convex usage metering into Redis and built the platform-usage dashboard with per-model and per-user agent-run breakdowns for billing and abuse detection. Agent runtime: a unified auth flow for Claude Code inside the sandbox (OAuth setup-token over PTY, token passed per run as env), project-level permission toggles, and GitHub repo linking that keeps the VM remote in sync. Containers that actually run Chromium: debugged crashpad / posix_spawn EAGAIN failures in Docker and standardized the image on node-slim.

Built AI agents for customer support and lead generation with LangChain + Superagent, served from a FastAPI backend. Created a voice modulator that clones a user's tone for video replies. Automated daily Facebook content across 30+ pages with Redis-backed pipelines, cutting manual work by ~90%.

Rebuilt the core web app frontend for faster loads and a smoother UX. Set up CI/CD with GitHub Actions + Docker for zero-downtime deploys. Migrated the backend from SQLite to Supabase (about 3x faster queries) and ran targeted email campaigns to 15k+ users.

Designed and deployed 10+ microservices for Gen AI workloads with Docker + Redis. Fine-tuned invoice parsers that turn scanned invoices into structured JSON at ~95% accuracy. Integrated NVIDIA RIVA for real-time call transcription and synthesis; built Stable Diffusion + ControlNet tooling for marketing images.

Contributed to SymPy's core and mechanics modules: a trigonometric rewrite rule for Bessel J, the orient_dcm / orient_explicit API in the mechanics reference-frame code, and correctness fixes to Mod, ceiling and expression printing.