Projects & Experiments

Whether a local agent loop (worker + evaluator) with ChromaDB-backed RAG and per-project memory could serve as a genuine productivity tool, not just a demo.

Built a working multi-mode agent with full-file RAG retrieval, SQLite checkpoints, and LangSmith tracing. Validated the worker+evaluator pattern at scale.

That metadata and governance are the true foundation for effective AI. Without good data definitions, the agent guesses wrong. With them, it reliably translates natural language to accurate SQL.

Validated the pattern end-to-end. The agent uses tools to inspect schema, read the dictionary, and run SQL in a loop — no hardcoded queries.

How an LLM chooses and invokes tools, receives results, and continues reasoning — without relying on framework abstractions.

Built a custom agent loop with tool registry, long-term memory, planner, and worker from first principles. Useful reference for understanding what frameworks abstract away.

The LLM-as-judge pattern for output quality control — a second model evaluates every reply and triggers a retry with feedback if it fails.

Deployed to Hugging Face Spaces. The evaluator pattern significantly improves response consistency and prevents hallucinations.

A scalable full-stack foundation: React SPA + FastAPI + Postgres (pgvector) + Redis + MinIO — learning each layer step by step.

Core infrastructure operational via Docker Compose. API, workers, and web UI scaffolded. Validates the architecture and data model before full implementation.

Whether meeting knowledge could be made persistent, searchable, and temporally aware — tracking how decisions and ideas evolve across meetings — without cloud dependency or an LLM API.

Built a full ingestion-to-retrieval pipeline: topic segmentation, structured extraction, temporal supersession linking, and a three-layer store (SQLite + JSON knowledge graph + lexical vector index). Works entirely offline with an optional LLM upgrade path.

Whether 17 coordinated LLM agents working in parallel — each owning a distinct role — could reliably turn a single raw idea into a complete, production-ready business case and product plan.

Built a full multi-tenant platform: LangGraph orchestration, FastAPI backend, Stripe billing, BYOK key vault, Jinja2 web UI, and an MCP server for IDE integration. Pipeline runs end-to-end with durable interrupt/resume at every human checkpoint.

A complete data lake — raw zone, AI ETL, structured serving, data catalog, and interactive analytics — can be built entirely on free Google services with no dedicated database, no data warehouse, and no cost per query.

Proved the concept: DuckDB WASM runs SQL in the browser querying data from Google Sheets, GPT-4o-mini extracts structured JSON from PDF/XML invoices, and the entire stack costs under $5/month vs $50–$1,000+ for AWS or Snowflake equivalents.

Enterprise architecture principles — governed data, controlled calculation engines, workflow automation — can be applied in low-code environments.

Production-deployed system handling multiple contract types, quality-based payout thresholds, advance payments, PDF generation, and automated email distribution.

How local diffusion models work in practice: download, memory requirements, precision issues on Apple Silicon, and pipeline differences.

Working t2i, i2i, and inpainting pipelines. Key lesson: MPS requires float32; attention_slicing helps on limited VRAM.