I build reliable AI systems where model behavior meets browsers, streaming protocols, storage, compilers, and experimental evidence. My work favors bounded inputs, deterministic replay, independent verification, and explicit claim limits.

This profile is a fast path through seven current systems projects. Each summary below names the public evidence that exists today and the conclusion that evidence does not support.

Curated portfolio navigation. The strict portfolio/projects.v1.json source binds every card to a reviewed default-branch commit, one public evidence surface, and one explicit non-goal. It is not a benchmark scorecard or live remote-state attestation.

Reproduce the SVG and verify its exact two-file bundle:

python3 tools/render_portfolio_map.py --check

The adjacent manifest binds the contract, renderer, semantic digest, seven immutable refs, and SVG bytes.

Task-aware browser-change evidence for asking whether a visual change breaks a user workflow or accessibility surface. The current authoring bundle contains two synthetic applications, four workflows, and 12 real deterministic Chromium checkpoints with accessibility, layout, action, provenance, and manifest records. It contains no official pair, released dataset, trained model, benchmark result, or accuracy claim.

A Linux loopback gateway for bounded multi-tenant Chat Completions traffic, weighted-deficit admission, validated buffered/SSE relay, cancellation, and signal-owned shutdown. Public evidence covers one controlled loopback workflow and one fixed-seed 28-job saturation run whose dispatches match an independent bounded oracle. It does not report throughput, latency, RSS, a fairness score, or a service-share benchmark.

A model-agnostic laboratory for verified out-of-core expert storage, bounded DRAM caching, cache-policy research, and a narrow BF16/AVX2 kernel. The closed M1-M4 bundle exposes captured command output, raw evidence, source-bound visuals, and explicit milestone reviews. Its M3 evidence measures synthetic modeled traffic, and M4 covers fixed synthetic GEMV batches on one recorded host; neither is an end-to-end inference or serving-speed claim.

A dependency-free static f32 compiler/runtime with checked graphs, explicit rewrites, reverse-verified arena planning, independently reconstructed execution plans, guarded allocation-free sessions, and a separate reference interpreter. Source-bound release-CLI and visual evidence now exercises three compiled-in workloads: the channel-affine slice proves mul_broadcast_f32 -> add_broadcast_f32 with 6/6 raw output words independently matched, while the six-step ReGLU slice remains separately captured. The project makes no benchmark, general-model, importer, or full-transformer claim.

An open-set keyword-spotting research system built around the failure mode that clip accuracy misses: false activations on continuous unrelated speech. Experiments 000 and 001 provide the measured linear baseline and development replay; all 1,001 registered thresholds failed the joint retention and false-event gates. Experiments 002-006 are retained engineering and incident evidence only: they produced no valid neural metric, reusable checkpoint, ONNX result, or continuous-replay score.

A clean-room lab for structural expert compression without silently relabeling dtype changes as compression. Its current M1 result is a pinned, bounded official-metadata target genome with raw records, a deliberate rejection path, and an 11-file visual atlas. No full checkpoint was downloaded or run, so M1 makes no compression-ratio, quality, throughput, active-compute, or state-of-the-art claim.

A fail-closed local preflight for PEFT LoRA checkpoints that inventories components, parses pinned configuration and safetensors headers, and emits deterministic structural evidence without importing model code or reading tensor payload bytes. Eight real CLI cases expose the current 8-of-17 rule slice. Even a clean run remains UNKNOWN; it is not proof that an adapter can load against a base model.

• Netveil · Python · Privacy and supply-chain security — offline pseudonymized audit receipts, guarded wheel execution, a published v0.3.0 evidence bundle, and explicit disclosure limits.
• K2DO · Python · Agent orchestration — a provenance-explicit nanobot derivative with routed DeepThink, real MCP subprocess fault labs, bounded cancellation, and three source-bound visual evidence suites.
• MeasureTrace · Python · Exact computing — rational m/km/mi conversion, independently recomputed receipts, reproducible packages, and real responsive browser captures.
• ShardLift · Python / PyTorch · Distributed training — crash-consistent checkpoints, deterministic recovery, and auditable real-SIGKILL evidence.
• KVCrucible · Rust / Python · LLM inference reliability — an offline conformance lab for unreliable KV-cache event streams, explicit uncertainty, replayable witnesses, and fault-injection evidence.

Also maintained: RecallLedger / note · UnitSentinel / units · Casefold / case · PasswordGenerator · GWorker · WitnessGap · A1220

• State the trust boundary and non-goal before making a claim.
• Bound bytes, shapes, work, queues, retained state, and diagnostics.
• Prefer deterministic state machines, property tests, pinned toolchains, and machine-readable evidence.
• Keep optimized paths answerable to a separate reference or verifier.