Work

I pool GPUs from different vendors and generations into a single inference backend of roughly 20 GB over Vulkan. FreeToken, a bleeding-edge MoE inference server, fronts the cluster and handles the elastic part: only the experts a token actually routes to need to be resident, so models with far more total parameters than the cluster's VRAM can still be served. My own scripts handle KV caching and system-prompt compaction, which is what makes the setup usable as a backend for coding agents.

The screenshot is a single node: a 35B-parameter hybrid MoE/Mamba model at NVFP4, running in 12 GB of VRAM with an 89K-token KV cache, the MoE expert cache and the SSM state slots all sized live against the memory budget.

The same rig trains as well as it serves: I've fine-tuned Qwen small language models with PEFT/TRL and trained diffusion image models. It also hosts my personal web assets, including this site and the services on the status page. I voluntarily maintain production-grade systems at every level of the stack so I'll survive the Big Tech apocalypse.

I compared the maintainability of human-written code against code from five AI coding agents (Claude Code, Copilot, Cursor, Devin, and OpenAI Codex) across more than 85,000 code definitions, scoring each with mature code-smell detection algorithms and comparing the populations with non-parametric statistics.

The finding: AI code carries 3–4x more structural debt, and it is the kind of debt reviewers don't see. The paper recommends automated quality gates on AI-generated pull requests rather than relying on human review alone.

During a co-op with Dow Chemical's Global Reactive Chemicals team, I independently built a Flask application that automates calorimetry analysis. It encodes adiabatic correction and exotherm-detection logic across 9 test types, and alerts on unusual or dangerous results from a wide range of testing environments and equipment.

The core design decision was an object-model abstraction that decouples each test type's parser from report generation. Adding a new instrument means writing a parser, not a new tool, which is how one app replaced 4 legacy VBA tools. It became the interdepartmental data standard, saving more than 1,020 hours (about 0.65 FTE) a year.

For a century, cloud chamber tracks have been classified by eye: alphas are short and thick, betas long and wandering, muons long and straight. A single camera makes that worse. A track angled away from the lens is foreshortened, which biases any length-based rule, and the rule returns a hard label with no confidence exactly where it is least reliable.

• Apparatus. A diffusion cloud chamber imaged by three hardware-synchronized Raspberry Pi cameras, calibrated with Zhang's method in OpenCV and validated by triangulating an object of known length before any particle data is trusted.
• Data. A spike-in protocol: background-only sessions (cosmic muons, radon-daughter alphas), then sessions with unlicensed alpha and beta sources at fixed positions, which gives weak ground-truth labels without breaking the bank or breaking into the IAEA.
• Models. Three tiers on the same features: classical length/width thresholds, a supervised ML baseline, and a Bayesian classifier that propagates calibration and triangulation error into a calibrated posterior over particle type. Each tier runs on both single-camera 2D and triangulated 3D features to measure how much 3D reconstruction actually corrects the bias.
• Monitoring. Shewhart and CUSUM control charts on session-level statistics flag anomalous chamber conditions; the deliberate spike-in shifts double as a validation of the monitor itself.

The proposal is approved and construction starts this fall. Results will be posted here.