
AI Engineer
Hiring Plug · Feb 2026 – Present
- Architected a production FastAPI backend (~88 Python modules, 10+ API surfaces) powering HiringPlug's AI-driven recruiting platform, serving 200K+ candidate profiles, 50K+ job postings, and 100+ agencies.
- Migrated HiringPlug's AI HR assistant from a single-intent MongoDB Atlas RAG system locked to one query type per conversation, degrading into hallucination after 5-6 turns, no pagination/deduplication, ~55s response time to a multi-agent architecture on Qdrant with a query orchestrator that decomposes requests across job, candidate, and FAQ agents within a single conversation, eliminating hallucination, adding proper pagination, and cutting response time to under 20s (as low as 15s); also introduced candidate summary generation during ingestion, previously absent, for complete and consistent retrieval metadata.
- Sole engineer building CoffeeCrew, an internal AI recruiting platform serving 2-3M candidate profiles: designed a multi-agent assistant where a task-understanding agent routes queries to specialist agents (JD analysis, candidate matching), with prompt-injection detection and content moderation on all pasted/uploaded JDs, achieving ~90% accuracy against a test set of injection/inappropriate-content attempts.
- Engineered semantic caching for common queries and model routing (lightweight tasks to Claude Haiku, complex tasks to larger models) with automatic multi-provider failover (Anthropic → OpenAI → Gemini) for cost efficiency and uptime.
- Built a parallel batch-scoring service processing candidates in batches of 25 against Qdrant-backed embeddings in 15-30s per batch, generating an LLM-based fit score and reasoning per candidate for filtering; also built JD and candidate parsing services (~3s per resume).
- Built an AI outreach system conducting candidate communication across WhatsApp, email, and LinkedIn in channel-appropriate tone, achieving ~90% accuracy in understanding and correctly shortlisting candidate replies, managing the funnel through to interview scheduling; also built an AI meeting-transcript scoring service.
- Owned CI/CD (GitHub Actions → Docker → AWS ECR → EC2), full test coverage, and platform-wide RBAC/rate limiting across all CC APIs; built and deployed a Model Context Protocol (MCP) server exposing recruiter chat as a tool, integrated live with ChatGPT, Claude, and Cursor.



