High-volume training-data infrastructure
AI training needed reproducible evidence of how real software changed and behaved over time. I built Docker-based pipelines that replayed commit histories, executed repositories, and captured terminal and test output as structured examples.
OutcomeConverted roughly 2,000 repositories into 100 GB of structured training data in about 12 hours.
- Docker
- Python
- Automation
- CI/CD
Multimodal document processing
A large book corpus had to become reliable multimodal training material, not just extracted text. I created a concurrent PDF/EPUB pipeline producing page images, rich text, layout metadata, and bounding boxes, with pixel-level validation and selective retries for corrupted renders.
OutcomeProcessed a 200–250 GB book corpus while recovering failed pages individually instead of rerunning entire books.
- Python
- PDF/EPUB
- Computer Vision
- Concurrency
Tool-using AI agent pilot
A product pilot needed an AI agent that could safely act, remember context, and expose its work through a usable interface. I built the permission layer, multi-step tool orchestration, REST APIs, persistence, conversation state, tests, deployment, and SolidJS chat experience.
OutcomeDelivered the working end-to-end pilot in approximately two weeks.
- LLMs
- Tool Calling
- SolidJS
- REST APIs
Additional full-time improvements
Incremental QuickBooks sync.Cut data-fetch time on large accounts from 5–8 minutes to 1–2 minutes by syncing created, modified, and deleted entities.
Storage redesign.Reworked scan results around shared entities, IDs, and option sets, reducing representative output from 70–78 MB to approximately 15 MB.
