Software Engineer · LLM Evaluation & Alignment · Backend / RAG Systems

AI-focused software engineer building production systems with hands-on frontier-LLM evaluation experience. I've been contracted to Anthropic (via Revelo) to evaluate Claude Code's production PRs — generating the preference data behind RLHF and reward-model training — and I now author expert coding trajectories at AfterQuery that teach frontier models to reason.

I like working where alignment research meets real engineering: not just labeling data, but building the RAG platforms and evaluation harnesses that turn "is this answer good?" into a reproducible, regression-tested metric.

Hands-on experience shaping how frontier models learn — from generating the preference data behind RLHF to building the eval infrastructure that measures whether a model is actually right.

• Frontier-model evaluation @ Anthropic (via Revelo). Evaluated & ranked 50+ AI-generated PRs across real-world Python repos — scoring correctness, test coverage, and edge-case handling at an 80% first-pass acceptance rate — producing the preference labels that fed RLHF and reward-model training.
• Training-data authorship @ AfterQuery. Author expert coding trajectories (reference solutions + stepwise reasoning) across 12+ repositories, 4 languages, and 6+ problem types, plus agentic coding environments that teach models how an expert resolves multi-file problems.
• RAG evaluation harnesses. Build eval systems that measure hit rate, MRR, precision@k, and LLM-as-judge groundedness/relevance on auto-generated test sets — making answer quality a reproducible, regression-tested metric instead of a vibe.
• Applied LLM systems. Embeddings & vector retrieval (pgvector), Claude API & Gemini, LangChain, Socratic RAG tutoring, and token-by-token SSE streaming with source citations.

Shipping production backends with an eye for latency, resilience, and clean service boundaries.

• Performance. Rebuilt chat-history storage as a standalone Go microservice over MongoDB, cutting response latency 263ms → 41ms (~84%) as an independently scalable service with its own datastore.
• Resilience & caching. Integrated 3 external FX APIs behind a Redis write-through cache with scheduled refresh and provider failover — cutting redundant upstream calls ~80%.
• Event-driven systems. Shipped an event-driven price-drop alert feature (Firebase Cloud Messaging) on a product ranked #1 of 13 company-wide; comfortable with async workers (Celery), WebSockets, and SSE.
• Architecture & quality. Multi-tenant APIs, JWT/RBAC auth, SSRF-guarded ingestion, encrypted BYOK, Dockerized services, and CI/CD — always backed by real test suites (115- and 70+-test pytest suites on recent projects).

• 🔬 Authoring expert coding trajectories & agentic environments that train frontier LLMs — @ AfterQuery
• 🏗️ Building RAG systems and scalable backends with FastAPI, Go, and Next.js
• 📚 Head of Education — mentoring 40+ engineers in DSA, algorithms & system design
• 🌱 Going deeper on system design, cloud deployment workflows, and Go internals

AlgoMentor — Graph-Guided RAG Learning Platform Multi-tenant full-stack SaaS (FastAPI + Next.js, 46 endpoints / 19 Postgres tables) with top-12 pgvector retrieval over Gemini embeddings, async Celery + Redis ingestion, and token-by-token SSE streaming. Includes a RAG evaluation harness (hit rate, MRR, precision@k, LLM-as-judge) and a prerequisite-graph engine that recommends what to study next. Shipped with a 115-test suite and CI/CD to DigitalOcean.

SafeSight Analytics — Real-time Video Surveillance Platform Async FastAPI backend for real-time monitoring — REST + WebSocket alerting, a configurable rules engine, role-based access, and a DeepFace vision integration. Backed by a 70+ test pytest suite and Docker Compose.

Languages — Python · Go · TypeScript · JavaScript · C++ · Java Backend — FastAPI · Gin (Go) · Node.js · REST · WebSockets · SSE · Celery AI / LLM — RAG · Embeddings · pgvector · LangChain · Claude API · Gemini · RLHF & Evaluation Data / Infra — PostgreSQL · MongoDB · Redis · MySQL · Docker · GitHub Actions · Linux Frontend — React · Next.js · Tailwind CSS

Google-backed (A2SV) competitive programmer — 1,000+ problems solved across LeetCode & Codeforces, 26+ rated contests. I care about clean architecture, correctness, and code that's a pleasure for the next person to read.

Open to roles in AI/LLM engineering, alignment & evaluation, and backend systems.