Generative AI, Agentic AI & Full-Stack AI Development
A software engineering foundation applied to Generative AI. Each area below links to the real project or content behind it.
Full-Stack AI Engineering
Complete products, not notebooks. A Next.js and TypeScript multi-agent app with live streaming output, a SQLite data layer, Google sign-in, Drive and Gmail integrations, Word and PDF exports, 138 test files and CI/CD through GitHub Actions. And a FastAPI RAG backend with a React chat app, Google Drive sync, client-side PDF export, 44 pytest test functions and push-to-deploy through GitHub Actions, PM2 and Nginx.
RAG & Retrieval
Answers grounded in a team's own meetings. Transcripts are cleaned, chunked by speaker and embedded with all-MiniLM-L6-v2 into FAISS, alongside a custom BM25 keyword index. Every question is searched both ways and the rankings are fused with Reciprocal Rank Fusion, narrowed by project and date filters. A four-prompt slot-state flow rebuilds each follow-up into a standalone query, so multi-turn questions keep their context.
Generative AI Engineering
Prompts managed as content and editable without a deploy, structured outputs validated with Zod, models bound per agent across several providers, prompt-injection defences that fail closed, and on-brand image generation with brand rules enforced in code. On the document side, a JSON prompt library and two-stage and chained pipelines turn meeting transcripts into Executive Summaries, requirements documents, user stories and timelines in the same templated structure every time, with model tiers on Groq and JSON repair on the output.
Agentic AI
A framework-free multi-agent system already delivered to a client — an orchestrator routing to six specialist agents, with orchestrator–workers, chaining, parallel rounds, a reflection loop, memory and human-in-the-loop built by hand. Autonomous tool-using agents on agent frameworks (Claude Agent SDK, LangGraph) are the active next phase, not a finished capability.
AI Automation & Integrations
Manual, repeatable work turned into pipelines and connectors — a design request becomes printable artwork, garment mockups and a print-kit ZIP; an OAuth 2.1-secured MCP server lets Claude generate images with OpenAI's models; and a Google Drive folder of meeting transcripts stays searchable through an incremental sync that processes only new or updated files.