At SWSAM Solutions, I built an autonomous planner that turns user goals into scheduled plans, with deterministic code assembling the plan so the LLM can’t invent IDs or schedules. I also designed a deny-by-default, human-in-the-loop permission layer gating all 31 write tools.
I cut planning cost about 77% per plan by migrating the agent to Claude Sonnet, with every end-to-end test passing. I also built a LangGraph discovery agent with 17 parallel retrieval tools and added hybrid document search over Qdrant.
At Transdata, I fine-tuned Qwen with LoRA on a quantized model for a customer profiling system, cutting inference cost 80% while keeping over 95% profile accuracy. I also built a multimodal video RAG system that reached 78% MRR@10 retrieval quality.
At Quid Sol, I developed and deployed production machine-learning models for NLP and computer vision. I also built a RAG evaluation framework as an open-source project, with recall@k, MRR, nDCG and faithfulness checks.
