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.

Released as v0.1.0, this Linux loopback gateway provides 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 dependency-free analyzer for bounded RBAC graphs with deterministic shortest escalation witnesses, Tarjan cycle reporting, scoped witness cuts, and an independent Floyd–Warshall verifier that never imports the analyzer. Public evidence includes 25 tests across Node 22/24 plus real CLI, desktop, mobile, full-page, GIF, SVG, and JSON artifacts bound to pinned Chromium and source hashes. The Orion policy is synthetic; there is no live IAM, enforcement, benchmark, compliance, or globally sufficient remediation claim.

A dependency-free exact scheduler for dependency-aware single-day planning with a separate optimality verifier, canonical SHA-256 receipts, safe CLI bundles, and a loopback-only dashboard. Public evidence includes 32 tests across Python 3.11-3.14 plus real CLI, Chromium desktop/mobile, SVG, and GIF captures bound by a manifest. Exact mode is capped at nine serial tasks; it does not model multi-user calendars, and caller estimates remain assumptions.

• CausalFence · Python · Distributed-systems consistency — released v0.1.0 with five causal/session guarantees, deterministic bounded witnesses, a hash-chained receipt, independent Floyd–Warshall replay, 60 tests, and a 13-file source-bound evidence bundle; the synthetic trace does not certify a database, establish production correctness, measure performance, or prove global witness minimality.
• FeatureSeal · Python · MLOps / temporal data engineering — released v0.1.0 with deterministic event-time and availability-time joins, six primary leakage reasons, canonical hash-chain receipts, independent SQLite replay, 73 tests across Python 3.11-3.14, and a 13-file source-bound evidence bundle; the synthetic fixture does not certify a feature store, establish production leakage rates, validate upstream truth, or measure model quality.
• QuorumGrad · Python · Adversarial ML / federated learning — released v0.1.0 with fixed-point coordinate-wise trimmed aggregation, explicit rank witnesses, canonical hash-chain receipts, independent O(n²) pairwise-rank replay, 88 tests across Python 3.11-3.14, and a 13-file source-bound evidence bundle; the synthetic round does not identify attackers, guarantee convergence, preserve privacy, certify a federated-learning deployment, or establish universal poisoning resistance.
• IntentGate · JavaScript / Python · Agentic AI safety — released v0.1.0 with deterministic proposal admission, human manager and privacy approvals, tenant/TTL/replay/one-effect controls, an independently verified hash ledger, and a 13-file source-bound evidence bundle; the synthetic HR fixture is not a model-quality result, regulatory-compliance claim, production security certification, or third-party integration.
• RecallLedger · Python · AI memory and storage — released v0.1.0 with a tenant-isolated canonical event ledger, dependency-free installed CLI, deterministic lexical oracle, fail-closed FTS5 candidate audit, 100% branch coverage across Python 3.11/3.12, and a real six-file installed-wheel evidence bundle; this Linux/POSIX pre-alpha trusts caller-supplied tenant context and makes no authentication, persistent-index, semantic-model, or performance claim.
• StrataFold · Python · MoE compression research — a pinned official-metadata target genome, raw M1 records, a deliberate rejection path, and an 11-file visual atlas; no full checkpoint, compression-ratio, quality, throughput, or speed claim.
• OrthoDrift · Python · Multilingual retrieval robustness — released v0.1.0 with typed grapheme provenance, globally minimal supplied-edit proofs, exact artifact replay, a verified offline report, and a 13-file source-bound visual bundle; the synthetic Şəki fixture is not a broad benchmark, dense-retrieval result, or native-language review.
• SensorProof · Python · Fault-aware robotics / sensor fusion — released v0.1.0 with deterministic fixed-point fusion, independently replayable decision certificates, 38/38 abrupt fault observations rejected or quarantined, a 98.18% RMSE reduction against the identical-stream ungated baseline, and a 13-file source-bound evidence bundle; the synthetic step-fault scenario is not a slow-ramp, correlated-fault, real-vehicle, production-covariance, or safety-certification claim.
• COWBOT · Python · ML systems observability — released v0.1.0 with graph-informed replay, a retained counterexample, source-bound incident visuals, and a frozen-unrun 128-pair holdout protocol; no holdout result, causal proof, throughput claim, or production validation is presented.
• UrbanLens · Python · Geospatial systems — released v0.3.0 with a deterministic spherical k-d tree, full-scan oracle verification, canonical query receipts, seven source-derived visuals, and an attributed historical dataset; it is not routing, current geopolitics, or survey-grade geodesy.
• Casefold Observatory · Python · Unicode security — released v0.5.0 with bounded collision graphs, canonical replay receipts, hardened CLI and offline HTML publication, real CLI/Chromium evidence, and no claim of authentication, confusable detection, or universal policy correctness.
• Auditable Receipt Extractor · Python · Multimodal document AI — released v0.4.0 with pre-upload batch validation, typed provider output, exact offline replay, content-addressed provenance, and a provider-free synthetic evaluator; public evidence is not live-model accuracy or proof that a model ran.
• PEFTLint · Python · Model artifact tooling — pre-released v0.1.0.dev0 with fail-closed LoRA checkpoint admission, eight real CLI cases, and an explicit UNKNOWN compatibility boundary: only 8 of 17 load rules are implemented, so COMPATIBLE remains unreachable.
• 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: UnitSentinel / units · PasswordGenerator · GWorker · WitnessGap

• 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.