🔐 Consistency · money movement that is atomic and idempotent
🧠 Reliable LLM apps · retrieval, guardrails, retries, observability
⚡ Performance · caching, rate limiting, queues, real-time ranking
🏗️ Clean structure · systems that are easy to reason about

learn() → build() → break() → debug() → document() ↻

Still learning. Always building.

Hybrid RAG Intelligence Engine

• Hybrid retrieval: Pinecone vector search + BM25, merged with Reciprocal Rank Fusion
• Semantic query relevance gate to filter irrelevant queries before LLM calls
• Redis caching and per-user rate limiting
• Multi-layer guardrails: prompt-injection defense, topic scoping, context validation, confidence scoring
• LangChain, LangSmith, FastAPI, Docker, GitHub Actions

Concurrent, ledger-based banking API

• Event-sourced, double-entry ledger with append-only credit/debit records
• Balances derived by aggregation, not mutable state, to avoid race conditions
• Atomic transfers using MongoDB sessions/transactions + idempotency keys
• Transaction lifecycle: PENDING / COMPLETED / FAILED / REVERSED
• Node.js, Express, MongoDB, JWT, bcryptjs

Real-time ranking with Redis Sorted Sets

• Continuously updated rankings built on Redis Sorted Sets
• Part of my wider Redis learning: TTLs, hashes, queues, BullMQ, pub/sub

Natural language → editable architecture diagrams

• LLM produces structured diagram data, backend streams it over SSE, frontend renders an editable Excalidraw canvas
• React, FastAPI, PostgreSQL/SQLite