Public projects
Diode Spectrum
Research audit · Python · cross-model representations · reproducibility
Diode Spectrum asks whether the effective rank of a target model’s representations predicts how well another model can reconstruct them on held-out prompts, after accounting for metric behavior, sample size, model family, and representation width.
The public release contains a frozen design, executable synthetic experiments, strict schemas, resource-bounded extraction, and preregistered bidirectional sensitivity analyses. All six directed maps among Qwen, TinyLlama, and SmolLM2 beat a geometry-preserving shuffled baseline on every one of 20 held-out splits in each of two separately frozen 64-prompt fixtures. The two TinyLlama pairings were directionally asymmetric while Qwen and SmolLM2 were nearly symmetric. This descriptive result covers three checkpoints; the larger 16-model hypothesis remains untested.
Explore the interactive explanation · Use the plain-language glossary · Inspect the repository
STRATA Transport
Research infrastructure · Python · causal evaluation · survey statistics
STRATA Transport asks whether a stable representation of demographic identity causally transports into human-aligned subgroup response differences.
The public release is a synthetic-only foundation. It implements compositional metrics, crossed resampling, weighted survey distributions, deterministic partitions, typed manifests, and fail-closed result binding. It has 147 passing tests at 96% branch coverage and contains no empirical finding.
Read the project note · Inspect the repository
Selected production work
The underlying employer code and data are private. These are high-level descriptions of the architectural work.
Constrained analytics agent
Natural-language requests compile through a custom DSL, and the model never writes SQL directly. The grammar is enforced in code and limits which operations the model can request.
Clinical reasoning workflow
A production agent combines grounded retrieval, biomedical terminology tools, a knowledge graph, self-critique, and human escalation. The system supports more than 20 pharmaceutical brands and reduced audience-development time from roughly 20 hours to roughly 2 hours.
Agent evaluation and provenance
The evaluation program is being built to examine full trajectories and spans in addition to scoring the final response. Outputs carry citation objects with untouched source records, giving reviewers a direct path from claim to evidence.
Orchestration, then simplification
An MCP-first platform supported dynamic agent discovery and parallel execution. Production reliability findings later led us to move important paths back toward a single agent with deterministic workflows. The evaluation evidence showed which paths did not benefit from the more complex graph.