- Turned inventory volatility into sellable demand — architected a 99%-accurate, self-healing pre-order engine across 50K products that flips SKUs between buy-now and pre-order as stock moves, so merchants never hand-manage availability and customers can still buy when stock runs out; a serverless, webhook-driven reconciler corrects platform drift (architecture)
- Built the live-selling revenue path — one ZegoCloud WebRTC broadcast fanned out to six destinations, each platform isolated behind its own state machine so a failed destination drops out without killing the sale; drove a 15% increase in live-session revenue
- Built apptile-harness, an AI context layer that models our services as a knowledge graph instead of markdown — giving engineers and agents accurate, multi-hop service context that stays in sync with the codebase
- Shipped an in-product “Get Help” AI assistant in the live-streaming service that chats with merchants and debugs their session live — grounded in harness for product knowledge and wired to the client’s network requests, so it sees exactly which calls are failing for that user; built on the Vercel AI SDK for model routing and tool calling
- Automated support triage end-to-end — inbound issues are parsed, classified by area and severity, and filed as Linear tickets pre-assigned to the owning team, removing manual assignment and cutting time-to-assignment
- Built a citation-backed RAG layer over internal specs, changelogs, and integration docs — giving the ops team self-serve answers on current product capabilities and cutting the volume of “can we support X?” escalations into engineering
- Achieved <500ms latency for near real-time alerting using async event-driven architecture and optimized data pipelines
- Constructed Grafana dashboards tracking KPIs like block times and transaction fees for data-driven performance improvements
- Built a full semantic search pipeline over Reddit — scrape → embed → store → query — with zero keyword matching; each post is embedded via a self-hosted all-MiniLM-L6-v2 sidecar and stored in ChromaDB
- Wrote the embedding service from scratch as a FastAPI microservice in its own Docker container; Express backend calls it over HTTP, keeping the ML runtime fully isolated from the API layer
- Queries are embedded at request time and ranked by cosine similarity inside ChromaDB
- Real-time bidding and order matching platform using Redis for atomic concurrency — handles simultaneous bids without race conditions
- Planned migration to Go to explore channels, goroutines, and locking primitives at the systems level
- RLM-style recursive inference pipeline over 590 Bhagavad Gita verses — treats the text corpus as an external environment and recursively prompts the LLM over verse-level snippets
- Answers long-context queries beyond single-pass limits with chapter- and theme-aware multilingual retrieval
- Platform for mass cold-email outreach to newly funded startups' CEOs and CTOs
- Cut query time on a 70k-entry founder table by adding targeted indexes on join and filter columns
- Full-stack cross-platform expense management app with Expo SDK 54 and a Dockerized Node.js/Express backend, featuring JWT auth, group-based expense tracking with role-based access
- Real-time sync via Ably WebSockets across all connected clients
- Browser extension providing hints and visualizations for tracing program execution and debugging with 80% accuracy
- Won Most Voted Award at Makeathon5 with 15K+ impressions on Twitter
Valhalla (OpenStreetMap Routing Engine) — Implemented real-time API warning notifications using Sonner toasts across route, isochrone, and optimized route endpoints — solving a long-standing UX gap that a prior GSoC attempt had failed to address
Valhalla (OpenStreetMap Routing Engine) — Added multilingual map label support with language persistence via localStorage, enabling English/German/local script switching across multiple tile providers
Autumn (YC F24) — Added dashboard CTA with APIs, boosting engagement by 50%
Mail.Zero (YC 25) — Eliminated 100% of redundant API calls, reducing backend load
Bifrost (Maxim AI) — Added UI configuration for vector and log storage backends
Klavis AI — Built a Google Analytics MCP server integration
Three Flags and a Race — building a webhook-driven convergence engine on top of an eventually-consistent third-party commerce API
Sharding a Pre-Order Engine — why one big selling-plan group doesn't work, and the webhook echo race that reordered every write
Counters, Not Logs — likes and comments at live-stream scale: one atomic in-memory increment instead of a per-like write, and a lock-free interaction layer where every count comes from the write that produced it
- Languages
- JavaScript / TypeScript, C++, SQL
- Frameworks & Tools
- Next.js, React.js, Expo, Node.js, Express.js, Prisma, Prometheus, Grafana
- Databases
- PostgreSQL, MongoDB, Redis, Firebase, ChromaDB
- Cloud & DevOps
- AWS, Docker, Kubernetes