Problem
Personal knowledge vaults and raw Markdown documentation contain unstructured notes and inconsistent linkages. AI agents cannot reliably ingest this raw context without risking hallucinations, missing topic boundaries, or processing private information.
What I built
An autonomous agentic pipeline that inspects Obsidian notes and documentation, extracts domain entities, and packages the content into structured Open Knowledge Format (OKF) bundles designed for safe downstream agent consumption.
Engineering
- Orchestrated multi-step autonomous knowledge structuring with Google ADK 2.0 and Gemini API.
- Implemented human-in-the-loop review-and-approval gates to control note boundaries and review sensitive content before packaging.
- Built an automated evaluation harness with programmatic unit assertions and LLM-as-a-judge scoring (Gemini Flash Lite) to assess extraction consistency and quality.
Current state
Core structuring pipeline, OKF bundle schema packaging, human review gate workflow, and the LLM-as-a-judge evaluation harness are implemented and validated.
Technology
- Google ADK 2.0
- Gemini API
- Python
- LLM Agents
- LLM-as-a-Judge
- OKF
Problem
Analyzing 70+ years of Formula 1 telemetry and race history requires managing schema drift, complex driver and constructor team changes across eras, and multi-table queries that LLMs cannot accurately generate without structured data models and a standardized interface.
What I built
An AWS lakehouse with a Model Context Protocol (MCP) server that allows AI agents to inspect curated analytical schemas and execute historical queries across Formula 1 telemetry.
Engineering
- Built an AWS lakehouse on S3, AWS Glue, and Athena using a Medallion architecture (Bronze raw ingestion, Silver cleansed tables, Gold analytical marts) populated from the FastF1 API.
- Transformed analytics models with dbt (dbt-athena), modeling SCD Type 2 snapshots for deterministic historical tracking of driver and constructor transfers.
- Developed an HTTP-based FastMCP server exposing schema discovery and controlled Athena query execution tools for AI agents.
- Automated post-race ingestion pipelines and state caching via AWS Lambda, EventBridge, and DynamoDB.
Current state
Data lakehouse architecture (Bronze, Silver, Gold layers with dbt models) and the FastMCP server query interface are implemented and operational.
Technology
- FastMCP
- MCP
- AWS S3
- AWS Glue
- AWS Athena
- dbt
- AWS Lambda
- DynamoDB
- Python
- SQL