I design research-grounded tasks and containerized benchmarks across mathematics,
machine learning research, operations, and data engineering.

My work spans benchmark task development with Handshake AI and Turing, alongside verifier engineering, evaluation infrastructure, and open-source contributions.

I worked on Terminal-Bench 2 through Project Dynamo and Terminal-Bench 3 through Project Insight, developing tasks across Operations, Mathematics, Machine Learning Research, and Data Engineering. My focus is turning domain problems into executable challenges with clear instructions, reproducible environments, and outcomes that can be checked.

I work with Turing on Terminal-Bench tasks for Meta, including scientific-domain and mathematical benchmark development. This work brings together research-led problem selection, precise task specifications, reference solutions, verifier design, and reproducible execution.

My scientific and mathematics work emphasizes a substantive reasoning challenge: a compact problem whose solution requires the underlying mathematics, with independently checkable results. I develop and review task packages, examine solver trajectories, and use validation evidence to distinguish reasoning failures from verifier defects or infrastructure problems.

Across these domains, I connect task design with reference solutions, automated grading, containerized environments, and failure analysis. The public examples below show that workflow in practice.

I work from a difficult agent task to the evidence needed to trust its result. The projects below follow that same path: design the challenge, build the grading system, then harden the evaluator.

The path: design → calibrate → grade → analyze → harden

My public Terminal-Bench contributions include operational decisions and database reliability challenges with outcomes that can be checked directly. My pull requests to Harbor are listed below and refreshed automatically as their status changes.

3 pull requests · View all on Harbor

• #2115 · Plan claims staffing from an incomplete backlog · OPEN
• #1985 · Recover coupon substitution trees from delayed fulfillment · OPEN
• #1969 · Reconcile drifting replicas within a transfer budget · OPEN

A closer look at replica reconciliation.

I designed a database reliability task that asks an agent to recover exact record drift from compressed replica sketches, choose one retry, and stay within a strict transfer budget.

Validation: Docker validation passed  ·  Oracle 1.0  ·  No-op 0.0
Model trials: GPT-5.6 Sol with Terminus-2 via OpenRouter  ·  0/5

View the pull request →    Read the failure analysis

Once the challenge is calibrated, the next problem is making its grading repeatable.

Infinity Megatron — enterprise grading  ·  PRIVATE R&D

I built the enterprise grading layer for Infinity Megatron, a private AI-agent evaluation platform. The pipeline covers rubric-based artifact scoring, verifier audits, Docker calibration, Pass@k trials, mutation testing, and structured reports.

_{Private implementation  ·  The animation presents the workflow at a high level.}

A repeatable grading pipeline still depends on an evaluator that fails safely.

Gandalf the Grader — evaluator reliability  ·  OPEN-SOURCE ENGINEERING

I submitted targeted patches for cross-platform process launching, UTF-8-safe I/O, Python 3.12 CI, and safe handling of malformed trajectories and judge responses.

View the submitted patches →

Across these projects, I use one repeatable process:

1. Frame the capability. Define the behavior being tested and the exact completion condition.
2. Build from controlled state. Derive expected results from data held by the verifier.
3. Calibrate both sides. Prove the oracle passes and shortcut or no-op solutions fail.
4. Protect the evaluation. Isolate agent code from fixtures, answers, and rewards.
5. Study and harden. Use trajectories to separate capability gaps from task defects, then improve the evaluator.

The public part of this work continues through benchmark and evaluator pull requests.