I build reliable AI systems across multimodal evaluation, browser automation, compiler/runtime infrastructure, inference reliability, and model tooling. I am especially interested in failures at system boundaries: unreliable event streams, model-artifact mismatches, unverified storage assumptions, ambiguous evidence, and real inputs that clean benchmarks miss.

My projects turn those failures into bounded, reproducible experiments with explicit trust assumptions and machine-checkable results.

ImpactDiff is a research lab for task-aware visual regression: whether a UI change breaks a user task or damages accessibility, rather than merely changing pixels. It captures canonical screenshots, accessibility trees, and bounded layout graphs around a real coordinate-driven task, while storing model-visible evidence separately from intervention metadata, execution traces, and labels.

The current milestone runs baseline and mutated candidates in distinct fresh BrowserContexts under one verified Chromium environment, compiles and rolls back typed palette and pointer interventions, derives and replays complete evidence records, and atomically publishes immutable visible/sealed release pairs. The Node 22 and 24 CI lanes exercise capture, provenance, recovery, and publication boundaries. A multi-application dataset, isolated feature runner, trained models, and benchmark result remain the next research milestone; the repository makes no accuracy claim today.

SSemaphore is a correctness-first local gateway for bounded, multi-tenant Chat Completions traffic. Its admission path combines estimated-service accounting with work-deficit round robin scheduling, explicit queue and body budgets, disconnect-aware cancellation, and a fixed upstream transport boundary.

The runnable gateway accepts a strict versioned policy from a private Linux file, consumes bearer credentials through one-shot environment references, and binds only to an exact numeric loopback address. End-to-end wire tests cover admission, authenticated forwarding, signal-owned shutdown, and listener cleanup; the release gate also exercises the race detector and 32-bit checked arithmetic.

Its bounded SSE relay validates one data: event at a time under total, per-event, count, idle-read, event, and total deadlines. It flushes validated chunks with bounded read-ahead, keeps the admission permit through a clean [DONE] and EOF, and separates pre-commit gateway errors from post-commit stream truncation. A real loopback test proves that the first chunk reaches the client before the upstream response completes.

ShardLift is a Linux crash-consistency lab for sharded training state. Its killer-demo kills an isolated checkpoint publisher with a real SIGKILL, recovers from verified immutable commit records, continues the recovered four-worker state, reshards it offline from four workers to three, and compares the result with an uninterrupted trajectory using logical-state roots and exact fixed-probe logits.

The canonical CPython 3.14 evidence lane runs both public crash barriers from a hash-locked wheel and PyTorch CPU runtime, re-verifies the retained bundles, and publishes the evidence and runtime provenance. Separate Python 3.11–3.14 lanes exercise the core package, typing, coverage, and clean wheel install. The claim is deliberately limited to process crashes on the recorded environment; it is not a power-loss durability or production-training claim.

TensorKiln is a dependency-free compiler for static f32 tensor graphs. Its current vertical slice includes checked rank 0-4 types, transactional graph construction, batched MatMul and broadcasting, an independent reference interpreter, dead-code elimination, and exact structural canonicalization with composable provenance.

The released alpha's graph-to-arena boundary derives compute lifetimes, applies a deterministic 64-byte interval planner, and returns a storage projection only after an independently coded reverse verifier reconstructs and agrees with every mapping, request, statistic, and allocation.

The merged execution milestone builds a minimal kernel/placement candidate and independently reconstructs its operands, dense layouts, storage, lifetimes, limits, and work before execution. Its arena session runs five explicit kernels, rejects invalid floating-point environments, guards each non-empty workspace, and can audit each kernel's exact write set. The first and repeated run remain allocation-free under wrapped C/C++ allocators. A replayable 128-DAG differential corpus, sanitizers, and a real ELF32/i386 x87 gate exercise the boundary; the milestone makes no benchmark claim.

KVCrucible is an offline conformance lab for reconstructing LLM KV-cache metadata from unreliable publisher streams. Its bounded Rust core validates canonical JSONL and tracks per-publisher Exact / Recovering / Unknown state. It materializes deterministic drop, duplicate, and reorder schedules, then compares fresh pristine and faulted folds without turning incomplete evidence into a false divergence claim.

The implementation includes opaque semantic fingerprints, atomic cache-view projection, property-tested fault materialization, a first-class per-stream convergence oracle, and a statically linked musl release gate.

peftlint moves LoRA adapter failures into a bounded preflight step without importing model code or allocating model tensors. The current vertical slice ships pure parsers for pinned PEFT adapter configurations and safetensors manifests, including duplicate-key rejection, explicit byte and JSON budgets, checked tensor-shape arithmetic, span validation, and payload-coverage proofs.

The merged inventory milestone turns those immutable manifests into exact, path-scoped evidence for saved LoRA pair closure, dimensions, and configured rank. It distinguishes modeled two-dimensional weight pairs from higher-rank convolution candidates, which remain Unknown instead of being mislabeled as incompatible. The evaluator imports neither PEFT nor Torch, reads no tensor payload, and never executes a retained rank-pattern regex.

The versioned compatibility contract keeps these findings separate from a full load verdict: base topology, selection, dtype, saved-module, and vocabulary evidence remain explicit future dependencies.

OrthoDrift finds minimal grapheme-level changes that make a retriever lose the relevant document. Its first research lane is Azerbaijani retrieval, while its core text and provenance machinery remains language-agnostic.

The current implementation runs deterministic BM25 experiments, reduces rank failures to minimal counterexamples, and emits self-contained JSONL evidence that can be independently replayed and verified against its recorded runtime environment.

• State the trust boundary, uncertainty, and non-goals before making a claim.
• Bound untrusted input, retained state, and diagnostic work before execution.
• Prefer deterministic state machines, property tests, golden vectors, pinned toolchains, and machine-readable evidence over screenshot-only demos.