AI / QA / AUTOMATION ENGINEER
I build AI systems, QA infrastructure, and automation workflows that prove they work.
Tested automation systems, AI workflows, eval harnesses, and QA pipelines — for teams that need reliable software, not fragile demos.
GATE: VERIFICATION IN PROGRESS
THE 30-SECOND SCAN
01 — THE 30-SECOND SCAN
SIX SYSTEMS. PICK YOUR PROBLEM.
E2E TEST FLEETS
AI EVALS & GUARDRAILS
OPS AUTOMATION
AGENT / MCP SYSTEMS
DATA PIPELINES
PROOF, NOT PERCENTAGES
02 — PROOF, NOT PERCENTAGES
No invented time savings. Every category links to an inspectable artifact.
AI QUALITYRAG evaluation harnessesInspect the eval output →BROWSER QAPlaywright E2E flowsSee the test report →OPERATIONSAutomation workflowsRead the run logs →RELEASE SAFETYRelease readiness gatesOpen the scorecard →HUMAN-IN-THE-LOOPAI assistant workflowsTrace the approval loop →DOCUMENTATIONRunbooks & workflow mapsReview the runbook →
02.5 — SYSTEM ANATOMY
Anatomy of a trustworthy automation
Select a stage — each links to the case study that proves it.
01 / TRIGGER
A defined input starts the run — a webhook, a schedule, a message. Never a person remembering to. If it can’t say what wakes it up, it isn’t an automation yet.
FEATURED CASE STUDIES
03 — FEATURED CASE STUDIES
BUILT, TESTED, INSTRUMENTED — AND HONEST ABOUT STATUS
Topology: an 85-runner fleet feeds a DAG orchestrator, whose runs are packaged by a hash-chain evidence signer (Ed25519-capable) and validated by the qa verify trust command. In parallel an autonomous SDET loop drives a 52-to-91-percent coverage ratchet; both paths terminate in the full-fleet CI gate.
01
Nexural QA OS — Signed-Evidence Runner Fleet
LIVE REPO + CIQA infrastructure for a 39-repo product federation.
- runners registered in the fleet
- 85
- measured line coverage
- 52→91%
- tests behind ratcheting thresholds
- 3,698
Topology: a 256-screen registry drives npm run verify, which fans out to Tier-1 static, Tier-2 logic, and Tier-3 engine-contract gates that all write into a machine-computed PROOF_REPORT. The same registry generates 257 Maestro flows feeding a Tier-4 device certification that honestly reads 0 of 256 until re-run.
02
Voza — 256-Screen Verification System
LOCAL PROOFVerification system for a 256-screen consumer language-learning app.
- screens green on Tier 1–3
- 256/256
- generated Maestro flows
- 257
- live service contracts wired
- 36/36
Topology: a 140-tool MCP surface backed by AST engines writes into an HMAC-anchored append-only proof ledger and proof graph, which feed the 68-of-68 release:check gates. In parallel a Python course auditor extracts 17k-plus claims into claim/source ledgers behind a scored gate (95 approved, 85 pilot, below 70 blocked) that reports into the same release check.
03
sage-kernel + Course Auditor
INTERNAL TOOLINGInternal SDLC platform + content-audit harness behind my own products.
- MCP tools served over stdio
- 140
- release gates green
- 68/68
- claims extracted at baseline
- 17k+
Topology: a session minter issues real JWTs to a headless Playwright driver that walks about 55 authenticated routes; a scorer checks console, network, and axe-core, and failures flow to triage-and-fix, which loops back into re-audit — three cycles drove the live dashboard from 92 to 99.6, guarded by a no-fake-data CI ratchet. The mock twin is explicitly excluded as an audit target.
04
Headless Dashboard Audit Loop
LIVE PRODUCTHardening loop for a live quant trading platform's ~55 production routes.
- route score average, 3 iterations
- 92→99.6
- authenticated production routes
- ~55
- console errors at final pass
- 0
Topology: commits flow through CI gates (about 1,374 tests, un-skipped by the packageManager fix) into an EAS cloud production build; Brotli log forensics reads build failures, an App Store Connect API key signs the submission with no Apple 2FA, and the app lands on TestFlight, exercising the live backend at api.joingiggl.app with production migrations 0048 to 0051.
05
GIGGL — Headless iOS Release Lane
LIVE BACKENDRelease pipeline for a consumer social iOS app.
- tests un-skipped by the CI fix
- ~1,374
- social surfaces wired to live backend
- 8
- production migrations shipped
- 0048–51
EVIDENCE LEDGER
04 — EVIDENCE LEDGER
CLAIM → ARTIFACT, LINE BY LINE
If a claim isn't on this ledger with an artifact behind it, it doesn't belong on this site.
14TOTAL CLAIMS
7T1
6T2
1T3
- 85-runner QA fleet with hash-sealed, redacted evidence bundles (Ed25519-capable signer) and a `qa verify` trust commandevidence/qa-run-* dirs + runner-registry.ts (repo nexural-qa-os)TypeScript · pnpm · sha256/Ed25519signed evidence bundlesT1
- Full-fleet E2E CI gate with SLSA provenance and artifact signingqa.yml · proof-gate.yml · slsa-provenance.yml · sign-artifacts.ymlGitHub Actionsgreen CI runsT1
- Found + fixed 11 false-pass "honesty bugs" — runners returning green while doing zero workrunner-conformance.test.ts + git log on feat/sdet-autonomyVitest · gitconformance suite + commit historyT1
- 256/256 Tier 1–3 verification on a 256-screen native app, machine-computed (never hand-edited)PROOF_REPORT.md via `npm run verify` (repo Voza-e2e-stabilization)TypeScript · screen registrygenerated proof reportT2
- 257 Maestro device flows auto-generated from the screen registrymobile/e2e/generated/ — 257 flows on diskMaestro · TypeScriptflow files on diskT2
- Tier-4 device cert 256/256 on iOS simulator (recorded 2026-06-14)cert runs on operator device — on-disk report honestly shows 0/256 until re-runMaestro · iOS Simulatorrecorded result — needs re-verificationT3
- 140-tool MCP SDLC server with an append-only, HMAC-anchored proof ledgerapps/mcp-server/tools.json + .sage-kernel/proof/ledger.jsonlNode.js · MCP (stdio)tool count + ledger on diskT2
- 68/68 release gates green with enforced coverage + complexity ratchetsscripts/release-check.mjs (123 test files)Node.jsterminal gate outputT2
- 23-course / 460+ lesson content audit baseline extracting 17k+ claimsAUDIT_BASELINE.md + audit-export.mjs (Supabase → markdown bridge)Python · pytest · Supabasegenerated ledgers + rankingT2
- Mechanical publish gate executed 160/160 labs — then the content-truth gate blocked an AI-generated 8-course batchaudit-courses.ts + PREMIUM_BUILD loop reportsTypeScript · Node.jsgate output + DO-NOT-PUBLISH verdictT2
- ~55 production dashboard routes driven 92 → 99.6 avg — 0 console errors, 0 network failures, axe-cleanlive production dashboard; audit harness was deliberately throwaway — rebuild before re-runningPlaywright · axe-core · JWT sessionslive prod routes (harness itself T3)T1
- No-fake-data CI ratchet bans `Math.random` / seeded mocks on cockpit screenscheck-cockpit-no-fake-data.sh + ci.ymlBash · GitHub ActionsCI script failing on an injected violationT1
- Exposed a false-green CI — missing `packageManager` silently skipped turbo gates; fix made ~1,374 tests actually runphase1-verify.yml · sage-gate.yml (repo giggl)Turborepo · GitHub Actionsbefore/after gate outputT1
- 20-page client site with a working lead-capture loop, live in productionweb-sage-ideas.vercel.app + admin lead cockpit with LIVE badge + SLAHTML/JS · Vercelpublic deploy + e2e walk-throughT1
- T1 LIVE REPO / CI REPORT / DEPLOYED DEMO
- T2 LOCAL PROOF / TERMINAL OUTPUT / DIAGRAM / RUNBOOK
- T3 WRITTEN EXPLANATION / PLANNED WORK
THE POSITION
Most automation portfolios are theater. This one is instrumented.
Every claim carries its artifact. Every gate generates a verdict. This site runs its own release gate in CI.
PRODUCTION SYSTEMS, NOT PORTFOLIO THEATER.
WORK SAMPLES
05 — WORK SAMPLES
REAL ARTIFACTS, PULLED FROM DISK
AI eval output
Answers scored against a rubric, not eyeballed. TS · eval script
Playwright browser proof
Critical flows run and verify themselves. Playwright · trace files
Release readiness report
"Ready" is a generated verdict, not a feeling. CI · readiness-report.json
Automation job log
Observable workflows — runs, retries, failures. Node · job runner
Architecture diagram
Real data flow — model, retrieval, approval, logging. built in code
Bug reproduction note
Defects with steps, expected vs actual, severity. QA · triage note
TECHNICAL WRITING
06 — TECHNICAL WRITING
How I test AI systems when the answer is probabilistic
You can't assert equality on a probabilistic answer — you assert properties.
PROVES: EVAL DESIGN · PUBLISHED →
What a reliable automation workflow needs before production
Not done until someone else can run, inspect, and recover it.
PROVES: OPERATIONAL MATURITY · DRAFT
How I think about Playwright coverage and flaky tests
Flaky tests are debt — fix the root cause or delete them.
PROVES: QA STRATEGY · PUBLISHED →
The difference between a demo bot and an operational AI workflow
A demo proves possibility. Operations prove repeatability under failure.
PROVES: PRODUCTION THINKING · DRAFT
A release gate is a trust contract, not a checklist
A gate earns its place when non-QA stakeholders can ship from it.
PROVES: COMMUNICATION · PUBLISHED →
THE OPERATOR
07 — THE OPERATOR
PRODUCTION SYSTEMS, NOT PORTFOLIO THEATER
I build production-grade automation: AI systems with eval harnesses, QA infrastructure that signs its own evidence, release gates that generate verdicts instead of feelings, and the pipelines that keep all of it observable and recoverable.
I run Sage Ideas, an AI-native studio shipping full-stack products and automation systems for operators who need software that works under pressure — and Nexural, a quant trading stack where a false green costs real money. That's where the no-invented-metrics discipline on this site comes from: my own systems only get to claim what their artifacts can prove.
SIGNAL
- mode
- founder-engineer
- focus
- automation + AI systems + QA infrastructure
- edge
- evals, release gates, pipelines, cloud
- stack
- TypeScript / Python · Next.js / FastAPI · Supabase / AWS
- standard
- production > prototype
HOW AN ENGAGEMENT RUNS
08 — HOW AN ENGAGEMENT RUNS
Every engagement runs the same five stages: audit (you get a findings doc), spec (a scoped plan), build (a working system), verify (a proof report), and handoff (a runbook plus a recorded walkthrough).
Five stages, five artifacts — every engagement ends with a proof report and a runbook, not a goodbye email.
SEE THE SERVICES →RESUME / CONTACT
09 — RESUME / CONTACT
If the proof holds up, let's talk.
Open to AI automation, QA engineering, SDET, test infrastructure, and workflow automation roles — or consulting projects: building an AI workflow, adding test coverage to a fragile product, creating evals for an AI feature, or turning a manual process into a monitored workflow.
Hiring for a full-time team instead? Same proof applies — grab the resume or email me.
FREE — SITE TEARDOWN
Paste your URL, get four Lighthouse scores and your six worst findings in ~20 seconds.
RUN THE FREE TEARDOWN →START A PROJECT — USUALLY THE FASTEST PATH
CONTRACT ENGAGEMENTS — FIXED SCOPE, PROOF INCLUDED
AI SYSTEMS
AI Workflow Build
You get a working AI workflow with an eval harness, guardrails, and an approval queue — plus the golden set and runbook that keep it honest after handoff.
BOOK A CALL →QA COVERAGE
Test Coverage Sprint
You get your critical flows under Playwright coverage, running in CI, with trace-backed reports your team can actually read.
BOOK A CALL →RELEASE SAFETY
Release Gate Setup
You get a ship/no-ship verdict on every build — checks, scores, and a readiness report generated by CI, not by vibes.
BOOK A CALL →AI / QA / Automation engineer focused on building tested AI workflows, browser automation, release readiness gates, and operational systems with inspectable proof.