Independent: Architecture Zero, TheStatic.tv, JobTrue · Remote

• Architected Architecture Zero, a white-label AI platform whose core is now open source under Apache-2.0, running live production instances from one core codebase: hybrid RAG (vector + custom Okapi BM25 + cross-encoder reranker), department-scoped knowledge bases with role-based access, multi-provider LLM routing (Claude / GPT / local Ollama) with failover, and local inference (7B to 30B-class models on an RTX 5090).
• Cleared a pre-registered, per-cohort eval gate on every row of one held pair (2026-10-24), its thresholds written before the numbers existed: answer correctness 88.3 tuned vs 76.6 on an outside-authored locked holdout (the 11.8-point overfit gap published), honesty 13/13, faithfulness 98.7, retrieval recall 90% (tuned) and 87% (full cohort), injection and hostile-web cohorts held 9 of 9. Answers and grades come from structurally different labs; the thin margins and the one rule change made before the run are reported beside the verdict.
• Built Eco Mode, a federated retrieval mesh: instances query each other's knowledge in parallel with per-caller scoped keys and fail-closed access boundaries. A public bot can never reach private departments; verified by adversarial testing as the least-privileged caller.
• Took a 36-finding outside security review of the whole multi-product workspace (8 High) to full closure across three rounds of outside re-checks: every finding verified at its cited line, every fix shipped with a test that fails on the old code, and the reviewer's final check re-ran those tests against the predecessor code. A pre-push read of my own found what the review had not: a production database's client rules let any signed-in account write its own balance and tier. Locked to server-only and audited for any sign of use.
• Shipped TheStatic.tv solo: a production Web3 streaming platform, 160+ authenticated API endpoints, GCP Livestream RTMP/HLS pipeline, a cron-driven billing engine with Redis distributed locks and viewer-scaled burn rates, and an ERC-721 contract on Ethereum mainnet.
• Shipped JobTrue (jobtrue.ai) solo: an AI job-search SaaS with Stripe billing, LLM fit-scoring with a server-side grounding layer that drops hallucinated resume items before storage, and per-user rate limiting. This resume is served from it.
• Operate everything built (hosting, CI/CD, monitoring, security review, documentation) plus nightly encrypted backups to immutable object storage (Cloudflare R2 bucket-lock + Google Drive) and an automated daily restore drill that boots each backup in a throwaway container: 9/9 checks, alarmed end-to-end via a fail-closed status probe.
• Calibrated the LLM-as-judge eval instrument against four independent references: a planted-error suite (51/51 caught across five rubrics, 2026-10-24), a human-adjudicated baseline (Cohen's kappa 0.61), a cross-lab re-grade (99.0% agreement, kappa 0.942), and an independently-authored harness (98.2% boundary agreement). Then audited my own locked exam, retired it for authorship bias, re-authored it via a third lab, and published the wider honest gap.
• Designed refuse-vs-fabricate honesty evals (cohorts demanding artifacts the corpus does not hold, scored by a purpose-built judge rubric with planted calibration), measured 12/12 and 13/13; drove an adversarial tier-isolation cohort from 3/11 to 11/11 with owner recall held at 100%. Built a four-arm cohort measuring over-refusal (the cost of a correct answer flipped into an unnecessary refusal), isolating it to 0/6 where it had been recorded as a wholesale refusal, and tracing that result to an uncontrolled test-fixture variable.
• Rolled the measured trust layer across every production instance, validating each port with like-for-like A/Bs on that instance's own cohort: +12 correctness on one peer, +18.4 recall with rank-1 hits doubled on another, and multi-turn follow-up retrieval driven 1/9 to 8/9 by a measure-gated resolver port.
• Eliminated hand-set status from the platform's trust dashboards: every number derives from stored eval runs or live runtime config at request time, never from a constant. Each reliability layer publishes either its live-derived measurement or an honest build status, so unmeasured layers stay visibly unmeasured rather than rounded up, and controls that are silent when disabled (rate limiting, injection scanning, egress screening) read their real runtime state instead of asserting "on". The tuned-vs-holdout overfit gap publishes as a first-class metric even when it widens.
• Root-caused silent production failures behind green health checks: an HNSW vector-index corruption (search dead while reads passed, shipped a knn tripwire plus per-run corpus fingerprinting) and two distinct vector-store data-loss mechanisms (flush-boundary loss on unclean stops; a file-watcher mass-purge during git deploys). Fixed and verified with back-to-back zero-loss deploy cycles.
• Open-sourced the platform's core under Apache-2.0 as a derivation rather than a fork: every file read in full and landed as a commit stating its keep-or-drop reasoning, on fresh history that structurally cannot leak what it never held. Gated on a six-lens adversarial pre-release review (27 findings, all remediated pre-publish) and a residue hunt for the class literal scrubbing cannot see (document shapes, date-subject pairs, metrics sampled from reality), which found 16 further items after a literal scrub verified clean.
• Exposed a private AI system to third-party clients read-only over MCP, then shipped an OAuth 2.1 authorization server so a hosted assistant could connect with a real handshake proven server-side. Deliberately zero write or confirm tools in the module, so the REST-only owner-confirmation rule keeps holding on a surface where a model holds the tools.
• Split every FastAPI entry point in the platform by job (9,825, 7,205, 4,856, 4,798 and 4,269 lines down to 400-650 each, nearly 600 route handlers into routers), every commit proven behavior-identical by snapshot diffs of the built app (each route's auth dependencies, the OpenAPI schema, startup order, middleware) and the full suite, with zero production incidents; then carried the method to the React clients under render snapshots. Found on the way a class of tests a refactor leaves passing while they test nothing, proven vacuous by mutation and re-pointed.
• Hardened every deployment, public open-source template first: stored credentials encrypted, MFA that fails closed, the caller's own password required at every write that creates durable authority, security state that survives restarts, production SSH closed to an identity-aware tunnel with deploys on pull-based runners, and untrusted document parsing moved into a locked no-network sandbox, its locks proven inside each live container. A cross-repository drift check holds the controls in place, itself fixed after four security fixes reached every private instance and missed the template.
• Built per-document access control for a RAG knowledge base: a Google Drive connector reads each file's real permissions and enforces them per document at retrieval, evaluated server-side in the vector store so a document the caller may not read is never fetched, never reranked, and never reaches the model's context. Demonstrated rather than asserted: a full administrator with no grant on a document retrieves it by neither the chat path nor the corpus browser, while the granted identity retrieves all of it.
• Built a platform's self-serve support layer as retrieval-optimized documentation (one self-contained answer per chunk-aligned section, troubleshooting entries keyed to the exact error strings the code emits so a pasted error retrieves its own explanation) and made supportability measurable: the shipped evaluation seed targets the help corpus with expected sources, so a fresh install scores its own onboarding and support answer quality on its first evaluation run.
• Orchestrated parallel AI coding agents through production backlogs, runs of 42, 60 and 84 agents, each agent in an isolated git worktree under a fact-checked brief and a fixed proof standard (a test that fails on the old code), and integrated every result myself: the diff read at its lines, the full suite and a browser harness of up to 2,000 checks after each landing. The 84-agent run put 361 commits into production; the 42-agent run, 260 commits through six gated deploys with zero regressions. Security defects the bug lists never named were found by the agents' own reads.
• Priced a cross-encoder reranker against its alternative with a cohort-matched A/B over 93 questions on the live corpus, adding the ordering metrics the harness lacked (nDCG@k, MRR) because recall cannot distinguish the answer at rank 1 from rank 5: the reranker bought +7.6 points of recall and +0.10 nDCG for +28.3 seconds per query, which kept it in the system and priced the hardware decision instead of guessing at it. A free read-only probe run first rejected the obvious cheap fix (cutting the candidate pool would have spent ~10 points of recall ceiling), so nothing shipped to learn it.

TEEMA · Everett, WA

• Maintain operational availability of testing workstations for aerospace manufacturing candidate assessments with a zero-downtime requirement.
• Serve as primary on-site contact for all technical issues; resolve hardware, peripheral, and connectivity problems and escalate to remote engineering when needed.
• Enforce strict site security and technical protocols as the on-the-ground liaison for off-site infrastructure managers.

Decentraland Foundation · Remote

• Provided technical support for creators and developers building on the Decentraland metaverse platform across Intercom, Discord, and Slack.
• Troubleshot SDK issues, scene deployment problems, wearables, and platform bugs; documented reproductions and coordinated with engineering teams for resolution.
• Leveraged 4+ years of hands-on Decentraland development experience to give expert-level guidance to the creator community.

Symetrix, Inc. · Mountlake Terrace, WA

• Drove live call answer rate from under 25% to over 80% by rebuilding NetSuite case-routing workflows and KPI tracking as primary NetSuite administrator.
• Cut Return Authorization turnaround from several weeks to 1-2 weeks by redesigning the process mid-2024.
• Hired as Tier 2 support; promoted to Team Lead in June 2023. Recognized publicly by the VP of Sales & Marketing for the call and case routing overhaul.
• Completed a NetSuite-ZOHO integration proof of concept; 2024 review: Exceeds Expectations across quality, initiative, technical skills, and dependability.

SAP Concur · Bellevue, WA

• Maintained 98% customer satisfaction across a high-volume enterprise Tier 3 queue for SAP Concur expense and travel products.
• Served as Comms lead on the Incident Response Team for P1/P2 issues. Owned customer and stakeholder communication under SLA pressure while Tech and Lead drove resolution.
• Queried client data with SQL to diagnose enterprise issues during Tier 3 troubleshooting; mentored team members and built performance dashboards.