Computational research engineer designing and building AI governance and reliability systems.
I build computational machinery for problems where getting an answer is not enough—you need a way to know whether the answer deserves to be believed. My work ranges from a hand-written C++ engine for mathematical research and a C++ geometry library with a hand-written foundation and later agent-assisted expansion, to larger systems whose architecture, interfaces, tests, and review gates I define before implementation, sometimes through coding-agent fleets.
Portfolio: harrisonwolf.github.io · Contact: harrisonwolf42@gmail.com · LinkedIn: harrison-wolf-5aa216187
A hand-written C++ engine for conjecture searches over pure Betti tables. On matched searches it runs approximately 70–115 times faster than Macaulay2 on the same machine. A larger search campaign surfaced counterexamples used in one of the three main theorems of Arithmetic in the Boij–Söderberg Cone. The repository includes benchmark methodology, raw data, and known limits.
A private neighborhood-intelligence system built around a versioned, eight-stage data pipeline, source provenance, fail-closed version checks, and a separate serving runtime. Coding agents produced most of the Python implementation under my architecture, schemas, scoring, and test gates.
The public Nexora Audit Edition is a synthetic, executable extraction of five reliability mechanisms: integrity-bound manifests, atomic publication, a journaled multi-target transition, typed SQLite carry-over, and evidence-qualified ranking. It includes tests, machine-readable claims, a source map, explicit limitations, and a correction record. It is released under AGPL-3.0-only for inspection and contains no product interface, real data, production configuration, or private history.
A local video-generation workflow built around two published open-weight backends on one 8 GB GPU. It combines a terminal interface, model orchestration, archived run settings and telemetry, seam/drift measurements, and a blind A/B comparison harness. It is an orchestration and measurement project, not a new video model.
A standard-library-only C++ geometry workbench. By the last checkpoint before agent-authored C++ entered the tree, I had written all 2,383 tracked C++ lines across 33 source and header files: primitives, polygon operations, collision and containment, an original ear-clipping triangulator, random polygon generation, drivers, and early visualization support. I later directed and reviewed coding-agent work that substantially revised the ear clipper and implemented the current Bowyer–Watson Delaunay and Voronoi layers, the custom TDD harness and much of its suite coverage, browser viewers, benchmarks, and terrain applications.
- Developing a private governance layer for multi-agent software work: preregistered checks, retained run evidence, adversarial review, and operator-reserved decisions. The design and implementation remain in progress.
- Evaluating frontier coding agents at DataAnnotation since August 2024: reasoning traces, code correctness, failure modes, containerized evaluation tasks, and quality review of other evaluators' work.
On compact problems, I keep the implementation small enough to inspect directly and measure boundary cases, performance, and failure. On larger systems, I define the architecture, interfaces, and acceptance tests before implementation, then review the result against them. The algebra core is hand-written; Planar’s mixed implementation history is documented in its project entry and repository. Coding agents produced most of Nexora's Python implementation under my architecture, schemas, scoring, and test gates. Each project identifies external models, algorithms, and implementation boundaries.
I also take on a small number of paid Agent Trust Reviews: bounded technical second opinions on one AI-agent architecture, evaluation, persistent failure, or deployment decision.