Ferrari racing driver. Amateur AI researcher.
Empirical work on inference economics, agent memory, and tool-calling reliability.
My newest public projects and research, newest first. I build tools for local AI, measure what actually works, and publish the evidence—including failures.
November 2026 · LLM evaluation
Does a model admit when the source does not contain the answer, or invent one? A live terminal observatory and reproducible benchmark for epistemic abstention, SQuAD v2.0 grounding, and TruthfulQA MC1. Includes resumable runs, paired comparisons, and a published stock-vs-WARMIND case study.
October 2026 · GPU and inference telemetry
A zero-build web dashboard combining GPU telemetry with LLM-serving metrics: VRAM, power, PCIe traffic, token throughput, latency, KV-cache usage, and inference costs. FastAPI, NVML, server-sent events, and a single-file frontend.
October 2026 · Independent inference research
A measured comparison of a two-GPU serving recipe against four-GPU configurations. Includes long-context retrieval, tuning A/B tests, a scheduler-stall reproduction, and checksummed evidence. Completed and cancelled experiments are clearly separated.
September 2026 · Model and runtime qualification
Independent evaluation of FP8 and NVFP4 checkpoints: capability retention, quantization fidelity, serving repairs, long-context retrieval, and sustained load. The report preserves failures and explains why the candidate was not cleared for production.
September 2026 · Linux hardware tooling
Reverse-engineered USB HID controls for EHOMEWEI portable monitors. Brightness unlock, HDR enablement, color troubleshooting, and automatic re-application on plug-in—without the vendor's Windows application.
September 2026 · Reproducible serving experiments
A field-validation harness and results archive covering decode and prefill, cache persistence, eviction correctness, scheduler fairness, and serving-stack qualification on a four-GPU Blackwell workstation.
- Measure, then claim. Preserve raw results and distinguish observations from explanations.
- Make comparisons reproducible. Record model, runtime, configuration, and hardware.
- Keep failures visible. A completed experiment is not automatically a production recommendation.
The Ferrari part is real. I race the 488 Challenge Evo. The AI part is also real — I just like the juxtaposition.



