AI Evaluation Engineer. Open to mid-level and senior IC roles, remote W2.

W2 full-time or contract-to-hire. Remote (US), based in Arizona. U.S.-authorized, no sponsorship.

171/177 adversarial prompts caught, synthetic corpus. Run it and compare.python3 build_corpus.py && python3 evaluate.py[171 / 177 adversarial caught]

Hire me - remote W2 Résumé (PDF)

01 / The receipt

$40,832.41 (synthetic corpus)

Run the command. The output below is what it prints.

python3 probe.pyCompounding error (loss): $40,832.40790

Re-checked 2026-10-06: survives

02 / What failed

What failed

Both error directions

171/177 caught 4/107 benign over-blocked

synthetic

6/177 (synthetic corpus)

python3 build_corpus.py && python3 evaluate.py

Prints: [6 adversarial missed]missed

Re-checked 2026-10-06: survives, caveat stated in the repo README

4/107 (synthetic corpus)

python3 build_corpus.py && python3 evaluate.py

Prints: [4 benign over-blocked]missed

Re-checked 2026-10-06: survives, caveat stated in the repo README

03 / Three cases

Three systems, two of them public. More on /work.

COBOL fixed-point truncation

cobol-pic-probe, public repo

Problem: PICTURE-clause fixed-point truncation compounds below the totals-reconciliation threshold, so a standard audit passes.
Built: A probe over a synthetic, fixed-seed payroll corpus: exact-decimal replay of PICTURE-clause truncation with per-field loss attribution.
Result: $40,832.41 synthetic loss across 47,312 synthetic records. Repeated runs return identical results.$ git clone https://github.com/mpuodziukas-labs/cobol-pic-probe && cd cobol-pic-probe && python3 probe.py
Limit: Synthetic payroll data, not production records.

Repository →

LLM adversarial gate

llm-adversarial-gate, public repo

Problem: An LLM pipeline with no adversarial gate ships wrong answers with no check and no replay trail.
Built: A rule-based BLOCK/ALLOW gate, mapped to five classes of the OWASP LLM Top 10, CI green.
Result: 171/177 adversarial prompts caught (blocked by the gate before release) and 4/107 benign prompts over-blocked (refused although benign), on a synthetic corpus (n=284) written for the repo.$ git clone https://github.com/mpuodziukas-labs/llm-adversarial-gate && cd llm-adversarial-gate && python3 build_corpus.py && python3 evaluate.py
Limit: Rules were tuned on this same corpus. There is no held-out set.

Repository →

Self-hosted inference

private fork of exo, ops-hardening branch

Problem: Managed-API inference means a single-vendor dependency and every query leaving the building.
Built: A private fork of exo (open-source distributed inference) on an Apple Silicon cluster serving open-weight models. Self-hosted first, cloud as fallback.
Result: An Apple Silicon cluster serving open-weight language models runs my own inference instead of a managed API; the primary keeps serving when the peer node drops.
Limit: Private, walkthrough on request. Losing the primary stops service.

Full proof record

04 / How the work is done

Checked by command, limits stated.

Every number on this site needs a public claims row to build; repo results print from a command in the public repos, and home-lab figures are walkthrough-only.
Each result names its corpus, and the corpus is synthetic.
Misses are published next to hits.
Each case states where it stops working.

Terms

W2 full-time or contract-to-hire.
Remote (US), based in Arizona. U.S.-authorized, no sponsorship.
Available now; start date set at offer.
I take live calls once a role is open.
Résumé (PDF)

05 / Work with me

Two of the three systems above are public code you can clone and run.

The third is a private fork. W2 full-time or contract-to-hire - Remote (US), based in Arizona - Available now; start date set at offer.

Hire me - remote W2