LLM training
A self-contained AI lab in one container
Built a single Docker image that turns raw Common Crawl web data into a tokenized training set with Spark, then uses NVIDIA NeMo (Megatron-LM) to pretrain a GPT from scratch and to fine-tune NVIDIA Nemotron. The same image serves a local LLM (Ollama), RAG over documents, and multimodal vision (CLIP, BLIP captioning, OCR, face analysis), fully offline against an enterprise data platform. A complete AI stack that runs with no internet.
Offline demo
Offline VAST Management System demo
A fully offline clone of the VAST Management System UI, built for sales. A web scraper crawls every link in the live console, captures each page, and stitches them back together under the same URLs as the real product, so navigation feels identical. It serves from a simple Python page with no backend, letting sales engineers walk customers through exactly how VMS works anywhere, even air-gapped. Thousands of lines across the capture and post-processing toolchain, spanning hundreds of routes.
AI agent
jAI, my own multi-model AI agent
A production agent I designed and run on three Cloudflare Workers. It streams answers live and calls 60+ tools: web research, images, voice and music, real PowerPoint, Word, Excel and PDF files, a Python sandbox, Microsoft 365 mail and calendar, and scheduled automations. In Consensus mode, OpenAI and Anthropic models debate with research tools until they agree, then one writes the final answer. Documents hundreds of pages long get retrieval plus a map-reduce summary that covers every page. Long-term memory, a per-token cost meter that compares providers, and an approval card on every external action round it out.
Founder build
The AI platform behind my own company
EcoLuxe is my beauty and wellness company. I shipped the first version of its AI platform in 30 days, solo, and have grown it to 17 cooperating serverless services: AI agents for Meta Ads, Amazon Ads, social marketing, creator screening and ambassador programs, with retrieval-augmented generation, multimodal vision, and human approval before anything goes live. About 80,000 lines of production code, guarded by 40,000 lines of tests and integrity checks on every deploy.
RAG
Grounded, multi-model retrieval on the edge
Runs entirely on Cloudflare Workers. Queries are embedded with text-embedding-3-small and matched against multiple Vectorize indexes (1536-dim, cosine) that keep product, advertising, and creator knowledge separate, with KV for shared state. Retrieved context grounds a multi-model step: a large reasoning model makes the decision while a smaller multimodal model screens images and audio, frame by frame. Outputs are structured and on-policy; anything low-confidence escalates to a human.
AI-assisted engineering
How I build: AI coding agents with guardrails
I build with AI coding agents and run them like an engineering team. Parallel agents lock each part of the codebase so they can't overwrite each other's work, a deploy is blocked unless the full test suite and integrity checks pass, live ad, payment and email accounts are read-only to the agents, and work orders mark every claim as verified, inferred, or unverified. The agents write much of the code; I set the architecture, the rules, and the bar for done.
Automation
50-state geo-distributed browser automation
Wrote a Playwright orchestrator that runs from all 50 U.S. state capitals via geo-targeting, performs human-like search and navigation, detects and engages a brand's owned properties, and captures per-city proof, concurrent, resilient, and self-documenting.
Enablement
AI keynotes, briefings & SDK workshops
As a Global AI Strategist I lead hands-on SDK workshops and keynote and executive briefings that democratize AI for enterprise customers and partners, meeting teams where they are and moving them forward.