Harsh Modi

About

Backend engineer who has shipped both high-throughput real-time systems and production LLM/RAG features. Built a database-free, local-first messaging service (Node/Express/Socket.IO) achieving ~1,500 msgs/sec at <25ms p95, and implemented a Go-based RAG recommendation pipeline with strict JSON/schema validation, catalog grounding, fallbacks, and eval loops that cut hallucinations to ~1–2% while reducing LLM costs ~60%.

Key Strengths

• Owned and shipped real-time messaging feature end-to-end (design → build → launch → iterate)
• Performance-focused backend engineering (handled ~1,500 msgs/sec; p95 latency <25ms; lookups 5–10ms)
• Pragmatic architecture tradeoffs (database-free local-first design with file storage + in-memory index)
• Robust lifecycle management for data without a DB (TTL indexing, scheduled cleanup, lazy deletion, restart rebuild)
• Production RAG/embeddings implementation with guardrails (catalog grounding, schema validation, fallbacks)
• Cost and latency optimization for LLM features (caching reduced LLM calls ~60%; responses ~200ms)
• Designed LLM evaluation loops (200–300 scenario test set, automated + human metrics, regression gates)
• Improved reliability via prompt constraints + validation (hallucination rate reduced from ~7% to ~1–2%)
• Built scalable ingestion/normalization pipeline for messy real-world data (validation, dedupe, synonym mapping, DLQ)

Experience

Education

Skills

Languages

Certifications