Personal R&D sandbox documenting the journey of building AI systems. Experiments, benchmarks, engineering deep dives, failures, lessons learned, and the occasional rabbit hole.
⚡ Active Stack & Tracks: Building Dali, an evidence layer for AI-assisted work. Current investigations focus on legal AI failures, citation verification, proposition support analysis, reproducible evidence records, verification drift, and failure taxonomies for high-stakes AI systems.
Evidence infrastructure for AI systems. Dali preserves source materials, verification outcomes, citation analysis, and reproducible evidence records so AI-assisted work can be inspected, reconstructed, and defended after the original interaction is gone.
Benchmarking local and hosted language models across latency, throughput, context utilization, and verification workloads.
Building a transformer from first principles to better understand the mechanics behind training, inference, attention, memory movement, and model behavior.
A structured dataset of legal AI failures mapped by what broke: fabricated citations, unsupported propositions, missing source trails, unverifiable outputs, privilege exposure, and workflow failures. Built to serve as a ground-truth corpus for evaluating whether AI systems preserve enough evidence to verify and reconstruct their work.