Chris Thom

• Architected and deployed multi-agent LLM systems using LangGraph + Autogen to automate internal tooling and user support workflows, reducing manual ticket resolution by 60%. • Engineered GPT-4o + Claude 3.5 hybrid pipelines with structured output parsing and tool calling for highaccuracy enterprise search and summarization. • Optimized RAG pipelines using Pinecone + LangChain for contextual retrieval across 10M+ document corpus, improving answer relevance by 35%. • Built real-time voice interfaces using Whisper + ElevenLabs integrated into mobile apps (React Native), enabling hands-free user interactions for accessibility features. • Leveraged MCP (Model Context Protocol) to dynamically route prompts between specialized models (Gemini for reasoning, Mistral for speed, GPT-4 for creativity). • Instrumented LangSmith for prompt versioning, A/B testing, and latency monitoring — reduced hallucination rate by 22% through iterative refinement. • Deployed vLLM + Unsloth for cost-efficient fine-tuning and inference of domain-specific LLMs on GCP Vertex AI.

• Engineered high-performance RAG architectures using LangChain + Pinecone (10M+ docs) and Weaviate + PostgreSQL — improving answer relevance by 35% and enabling semantic search with <200ms latency. Built hybrid keyword + vector search systems for enterprise applications. • Implemented and optimized production RAG pipelines using Pinecone, Weaviate, and FAISS — supporting contextual retrieval for LLMs across multimodal datasets (text, image metadata, user logs) at scale. • Implemented async Python workflows (asyncio, Celery) for concurrent model inference, media processing, and webhook handling — enabling scalable, non-blocking AI pipelines serving 10K+ requests/month. • Built real-time data ingestion and preprocessing pipelines for generative AI platforms — cleaning, deduplicating, and structuring user-generated content before vectorization. Integrated streaming via Twilio/Bland.ai for live conversational data.

• Developed video generation backend integrating OpenCV + Stable Diffusion Video pipelines for real estate virtual staging — processed 10K+ videos/month. • Implemented LangChain agents to automate video script → voiceover → asset eneration → publishing workflows — cut production time from 3 days to 4 hours. • Engineered GraphQL + Flask APIs to serve multimodal AI outputs (image, text, video) to OpenReact/React Native frontends with <1s latency. • Pioneered use of Ollama + LlamaIndex for local LLM inference in client demos — enabled offline-capable AI features without cloud dependency. • Collaborated with CV researchers (3 co-authored publications) on diffusion model optimization for edge deployment — achieved 3x speedup on mobile GPUs.