Francisco G.

Hands-On AI/ML Engineer | I Build Recommender Systems, LLM Apps & AI Agents (Claude, MCP) | Python, PyTorch, AWS | E-commerce, Streaming & Voice | NLP | PhD, ex-Amazon 👁‍🗨 Overview I build and ship production machine learning myself, end to end. No handoffs: I design it, write the code, measure the result, and stay on it until it is running. I bring 10+ years of hands-on ML at Amazon, Roku and Shopify, a PhD in machine learning from UMass Amherst, ten peer-reviewed papers and a granted US patent, and I apply all of it to LLMs and AI agents today. ✅ What I Build: • Recommendation systems: candidate generation, ranking, "customers also bought", homepage, search and email personalization, contextual bandits • LLM applications: RAG over your documents or catalogue, structured extraction from PDFs, classification and routing, semantic search, fine-tuning, an evaluation set from day one • AI agents and automation: Claude agents connected to your systems through MCP, read and write tools separated, human approval wherever money is involved • NLP and conversational AI: intent classification, entity extraction, product Q&A grounded in your data, semi-supervised learning when labels are scarce • ML engineering and strategy: architecture reviews, roadmaps, deployment, MLOps, A/B testing, a straight answer on what to build first 🎯 Recent Hands-On Builds: • Scaling candidate generation for a shopping app used by millions, bringing LLMs into recommendation where classical models miss product meaning • Built a store-wide recommendation engine for a fast-growing DTC brand and wrote its personalization roadmap • Researched, built and deployed ranking models serving millions of streaming households at Roku • Built a production semi-supervised NLU pipeline at Amazon Alexa from hundreds of millions of unlabelled utterances, published at NAACL 2021 • Built Claude agents that triage my team's email, invoices, tickets and pull requests through MCP, in daily use, a person on every approval • Delivered a read-only ML architecture review for a property-tax platform on Upwork in seven days, with a 12-month roadmap 💼 Industry Expertise: • E-commerce and marketplaces: recommenders, personalization, search ranking, catalogue automation • Streaming and media: ranking and recommendation for large catalogues and short sessions • Voice and conversational AI: natural language understanding at Alexa scale • Operations teams: document, email and approval work handled by agents, human in the loop 🔧 Technical Stack: AI/ML: Claude, Anthropic API, Claude Code, GPT, GPT-5, GPT-6, Codex, Gemini, open-weight models; RAG, embeddings, vector search, fine-tuning, semantic search Agents: MCP, tool use, structured outputs, prompt caching, model tiering, human-in-the-loop approval Recommenders: candidate generation, learning to rank, re-ranking, contextual bandits NLP: intent classification, entity recognition, semi-supervised learning (pseudo-labelling, distillation, VAT, CVT) ML: PyTorch, deep learning, reinforcement learning, meta-learning, ranking models Data and serving: Python, SQL, Snowflake, Snowpark, PySpark, FastAPI, Postgres, AWS Evals: labelled sets, regression suites, LLM-as-judge calibrated against people, online metrics, A/B testing 📊 Measured Impact: • Up to 15.25 percent relative error reduction on intent classification at Alexa scale (NAACL 2021) • Up to 1000x speedup for GPU path planning across hundreds of agents (ICRA 2014) • One correctness bug and one time-critical data gap found in a seven-day review • Every number here has a public source or a client who can confirm it 🎓 Credentials: • PhD and MS in Computer Science, UMass Amherst, reinforcement learning • Ten peer-reviewed papers: NeurIPS (first author), NAACL (co-first author), AAMAS, RLDM, ICRA, IROS, SCA, MIG, CAVW • US Patent 10,025,819, natural language to database queries, cited by 40 later patents • Reviewer for NeurIPS, ICML, ICLR, AAAI and IJCAI; GEM Fellow; internships at Microsoft Research and Adobe Research 🌟 What Sets Me Apart: I am research-grade and I still write the code. Most people building with LLMs can show you that an agent runs; I can tell you how often it is right, what it costs, and where it should stop and ask a human. The evaluation set is the product spec. For agents, autonomy is tiered by blast radius, and nothing that commits money is decided by something you cannot hold accountable. 🤝 Working With Me: You work directly with me and I do the work. Scoping first: what we are building, what "working" means, how we will measure it. Then milestones you can see, working code early, documentation at handover. Strategy calls, fixed-scope reviews and builds, or a part-time retainer. Boston, Eastern time. English and Spanish. Have an AI feature or a recommender that needs building or fixing? Send me the details and I will reply with exactly how I would build it.