Parmpal S.

I help businesses turn fragile AI prototypes into reliable, deterministic production systems—reducing hallucinations, cutting latency, and engineering scalable multi-agent architectures. I am a Lead AI Engineer with 9+ years in AI/ML and 4+ years architecting production LLM, RAG, and Agentic systems. I own delivery end-to-end: translating business logic into state graphs, building FastAPI/PostgreSQL backends, setting up automated evaluations, and deploying containerized services to the cloud. A capable model is only 20% of a successful AI product. I engineer the surrounding 80%—separating probabilistic LLM reasoning from deterministic application logic to handle messy data, concurrency, tool safety, and real-world edge cases. RECENT PRODUCTION SYSTEMS BUILT • Multi-Tenant Voice & Chat AI Receptionist (LangGraph & WebRTC): Architected a config-driven voice/chat agent platform using Pipecat, Deepgram STT/TTS, and PostgreSQL row-level security. Manages hundreds of live patient appointments daily, with double-booking blocked at the database level and real-time clinician triage. • Autonomous Finance Ops Agent (n8n, OCR & FastAPI): Built an inbox-to-ledger reconciliation agent processing bank receipts and invoices with ~90% extraction accuracy across 1,000 validated test emails, complete with risk scoring and human-in-the-loop audit trails. • High-Volume Talent & Career Platforms: Engineered an AI resume-parsing and matching pipeline processing ~30,000 resumes/month (GPT-4o, Claude 3.5, pgvector) and an autonomous career copilot receiving ~800 user profiles daily. • RAG & Policy Fine-Tuning: Built RAG pipelines on Elasticsearch and Pinecone, plus supervised fine-tuning of GPT-4o for complex policy interpretation achieving a 96% LLM-as-a-Judge pass rate. • AI Research & Simulation Engineering: Built the core codebase and ran experimental benchmarks for research teams publishing in U.S. DOE OSTI (LLM-guided Reinforcement Learning with RAG) and IEEE Xplore (distributed privacy-preserving edge deep learning). WHAT I CAN BUILD FOR YOU • Agentic Workflows & Custom MCPs: Deterministic LangGraph state machines, multi-agent orchestration, human-in-the-loop approval gates, and custom Model Context Protocol (MCP) servers. • RAG & Knowledge Engines: Vector and keyword retrieval, reranking, and automated faithfulness and hallucination scoring of answers against their source context. • Real-Time Voice & Multimodal AI: Low-latency streaming voice agents (WebRTC, Pipecat, Deepgram) and OCR document extraction pipelines. • Production AI Infrastructure: Async FastAPI services, PostgreSQL/pgvector, Redis, Docker, and observability (Langfuse, OpenTelemetry). HOW I WORK 1. Deterministic Boundaries: I never let an LLM guess where deterministic Python code or database constraints should enforce the rules. 2. Evaluation-Driven: Every pipeline ships with test datasets and automated accuracy, faithfulness, and cost tracking—no "vibe-checking" prompts. 3. End-to-End Ownership: You get one senior engineer who handles the system architecture, LLM orchestration, backend APIs, and cloud deployment. Whether you are scaling an MVP into production, fixing a hallucinating RAG pipeline, or automating complex operations, send me your current stack and biggest bottleneck. Let's map out the architecture.