What makes this different from staff augmentation?

One person with seven specializations unavailable in staff aug. COBOL PICTURE-clause forensic find in 127 minutes that 3 years of external audits missed. 3-5x lower cost for broader coverage.

Is the sovereign inference cluster operational?

Yes. 55 commits ahead of upstream on a production fork of exo (upstream lineage). RDMA ring topology. 389ms P50. AWS Bedrock and Azure OpenAI format-compatible. $26,400/mo cloud inference eliminated.

What audit-trail artifacts does the sovereign inference build produce?

SHA-256 hash-chained chain-of-custody on every inference output. Quality-management-system documentation, bias logs, and human-oversight records with a deterministic event-sourced audit trail, operational now.

What compensation are you targeting?

Negotiated privately, by role and scope — remote · W2 · skills-first.

100% remote. US-based, Mountain Time. Async-first. No relocation.

What kind of engineer are you?

Agentic architect and AI/data engineer — agentic systems, sovereign inference infrastructure, and adversarial-AI validation, every result replay-verifiable.

What proof do you offer?

Deterministic replay artifacts. SHA-256 hash-chained chain-of-custody. COBOL PICTURE-clause forensic audit in 127 minutes. 27× throughput. 96% latency cut. 0 false negatives across the public adversarial corpus (140/140 caught, n=220). All verifiable at puodziukas.dev/proof.

Why hire you instead of engaging a consulting firm?

Single operator, full-time, inside your perimeter. The person who scoped the work does the work — every output logged and replay-verifiable.

Yes. Standard MNDA on day one. All work product stays air-gapped.

What stacks do you work in?

COBOL · Rust · Python · TypeScript · Next.js · LLM APIs · AWS · GCP · Terraform · Pulumi · OpenTelemetry · NIST FIPS 203/204/205.

puodziukas.dev · remote · available Q3 2026

LLMOps / FMOps Engineer

Deterministic replay over a 47,312-record deterministic replay corpus. 0 divergence, hash-verified. LLM gate: public corpus n=220 (140 adversarial + 80 benign), 140/140 caught, 0/80 over-blocked.

a 47,312-record deterministic replay corpus, re-run on every model push. Zero false negatives across the public adversarial corpus (n=220): 140/140 caught, 0/80 over-blocked.

reproducible · verified · available Q3 2026

The problem

A sampling-based eval suite passes, and a hallucination still surfaces six weeks later in a client report. The gap is not in the eval suite itself: it is in what the eval suite was never measuring.

What exists, reproducible

a 47,312-record deterministic replay corpus catches the failure class your sampling-based evals miss. It runs in CI. Every model push gets the same adversarial probe. Zero false negatives across the public adversarial corpus (140/140 caught, 0/80 over-blocked, n=220), not a claim.

Relevant projects

→COBOL PICTURE-clause truncation surfaced via adversarial probe. Missed by standard monitoring.
→Deterministic replay pipeline over a 47,312-record deterministic replay corpus: zero false negatives across the public adversarial corpus (140/140 caught, 0/80 over-blocked, n=220)
→OWASP LLM Top 10 adversarial test suite: injection, jailbreak, hallucination, consistency
→Inference API cost: $26,400/mo → $0 via local inference mesh (self-hosted), quality parity
→Automated retrain trigger on drift detection, gated by the deterministic replay suite
→Prompt versioning architecture with A/B routing and statistical significance gate
→KV-Cache aggressive strategy: 40-60% token cost reduction documented

What you receive in 30 days

Adversarial eval suite deployed in your CI pipeline (Week 1)
Hallucination rate baseline via deterministic oracle, not sampled estimate
Prompt versioning + A/B routing with statistical significance gate

Companies at this intersection

Weights & Biases · LangChain · Arize AI · Scale AI · Comet ML · Fiddler · Helicone · Langfuse

Working terms

W2 remote · async-first · deliverable-gated output · first artifact in 48h · available Q3 2026

Work with me →

WRITTEN SCOPE FIRST · MILESTONE-GATED · REPLAY-VERIFIED

Full hire page →

All results under NDA. Representative figures. SHA-256 hash-chained audit trail available on request. · puodziukas.dev