PERSONAL PORTFOLIO / PUBLIC WORK
Eight public projects, grouped by the research problem they address and the evidence they expose.
Research infrastructure for reproducing low-level failures while preserving exact identity, provenance, and independently checkable evidence.
A simulation-first research platform for deterministic, reproducible Windows driver reliability experiments. It normalizes cases, resumes interrupted campaigns, minimizes exact-signature triggers, and produces independently verifiable evidence bundles.
Evidence surfaced Canonical Case IR, seeded scheduling, an append-only SQLite journal, exact signatures, 3/3 simulated replay, recorded G3, E1, E2, and minimization artifacts, semantic verification, and multi-platform CI.
Systems that make product risk, model change, and release decisions explicit and independently checkable.
A deterministic product-security decision layer for AI coding agents. It derives applicable controls from bounded repository evidence and records every decision in a Control Evaluation Ledger.
Evidence surfaced 16 risk packs, 116 controls, 9 composition rules, profile and mutation cases, real-repository evaluations, and multi-platform CI.
A vendor-neutral release-evidence layer for deployable AI models. It binds exact artifacts, paired evaluation, statistical policy, provenance, and regression onset into a portable Model Change Report.
Evidence surfaced Reproducible release records, explicit promotion policy, provenance, regression localization, and an independent verification path.
Systems that keep automation inside bounded tasks using inspectable evidence, explicit mutation authority, and reproducible intermediate artifacts.
Static analysis and a portable Agent Skill for AI-edited Python repositories. Weak findings stay review-only, while deletion requires a reviewed, hash-bound plan and isolated validation.
Evidence surfaced Import reachability, Git chronology, normalized AST evidence, 25 adversarial safety scenarios, real-repository smoke tests, and multi-platform CI.
An installable runtime, MCP server, and portable Agent Skill that gives coding agents persistent task, mutation, review, and completion boundaries. It binds verification to the current worktree, preserves project decisions, blocks unauthorized edits, and ends review at a deterministic SHIPPABLE gate.
Evidence surfaced 116 collected tests, Windows, macOS, and Linux CI across Python 3.10 and 3.13, seven versioned releases with checksums, deterministic boundary and mechanics benchmarks, and a published research evaluation plan. The historical agentic A/B result is retained but explicitly marked superseded.
Research tools for stress-testing novelty claims and turning incomplete progress into evidence-bounded advisor conversations.
An evidence-first Agent Skill for adversarial scholarly novelty audits. It freezes claims before retrieval, identifies Minimal Prior Sets, applies historical cutoffs, tests bridge evidence, and records what the search could not establish.
Evidence surfaced Claim maps, provider-normalized records, evidence-bound prior sets, temporal states, citation-graph bridges, deterministic report invariants, three public case-study classes, and exploratory measurements from 82 reviewer-annotated cases.
An evidence-grounded Agent Skill that turns raw research progress into a decision-ready advisor meeting. It separates observations from interpretation, carries unresolved requests forward, and frames one advisor-answerable decision without inventing completion.
Evidence surfaced Portable Research Meeting State validation, source and numeric grounding checks, 15 behavioral and adversarial case definitions, eight routing-collision cases, longitudinal holdout structure, and six-job cross-platform CI.
Applied computer-vision work for preserving weak visual signals and making uncertain recovery decisions inspectable.
A local-first computer-vision toolkit for recovering spatial structure carried by subtle color differences. It tests competing hypotheses, preserves auditable artifacts, and abstains when evidence is weak.
Evidence surfaced Masks, overlays, quality signals, provenance, uncertainty states, and algorithm notes.