How I build.
A real excerpt from one of my production systems — not a summary, the actual words:
“It fixes the failure mode we hit on #54: main got a new commit but no deploy fired (a dropped Actions event / skipped-or-failed deploy gate / a flake), leaving prod stale with no one watching.”
“WHAT IT CHECKS (honest scope) [...] the deploy JOB, not a /health 200 [...] This is a PROXY: it cannot see a Deploy run that reported success while production did not actually roll [...] on 2026-08-07 that left production 34m31s behind while every passive probe stayed green.”
Reusable, versioned instructions — for requirements and security, not only code.
113 authored skill directories, including skills that route between the others and audit anything adopted from outside.
CI that knows which checks can stop a release and which only inform it.
One workflow's own comment: a golden-set gate is hermetic and blocking; a latency baseline is advisory.
Tooling built from named incidents — a deploy-watchdog opens with the incident that caused it.
02 — Engineering operating model
One lifecycle, held end to end.
Eight stages I work across rather than hand off. Select a stage to read what it means in practice, which evidence supports it, and where it appears in the independent systems below.
Discover — find the customer problem
Customer conversations, UAT sessions and production complaints, read for the problem underneath the request — before anything gets designed.
Customer delivery, UAT, troubleshooting and stakeholder collaboration across Oracle Analytics, on Oracle Cloud Infrastructure and Oracle Analytics Cloud.
SaafSaans exists to answer one practical question directly: what does current air quality mean for my plans, and when might conditions improve?
- customer problem
- users
- outcomes
Frame — write the assumption down
Requirements, constraints, risks and data, turned into acceptance criteria specific enough to test — and non-goals named, not implied.
Aegis Contracts tags every assumption CONFIRMED (the owner’s input) or DEFAULT (change-controlled) in one file. Nothing is silently assumed.
- requirement
- acceptance criterion
- risk & ambiguity log
- non-goals
Architect — contracts before code
Systems compose when their boundaries are versioned. The artifact schema, the traceability model and the event catalogue get designed first, so tools underneath them can be replaced without renegotiating what anything means.
Architected AI-driven quality engineering frameworks across enterprise cloud services at Oracle.
Aegis Contracts — three versioned contracts: a canonical artifact schema, a traceability node/edge/coverage model, and a CloudEvents catalogue. Frozen at v0.1.2, additive-only within v1. Coverage is computed at the acceptance-criterion level, as a first-class object, not a spreadsheet column.
- requirement → criterion
- criterion → test
- test → execution
- defect → prod signal
Build — the AI workflow itself
The workflows, integrations and platform capabilities that turn a design into a working system — including the AI-assisted parts, which get the same discipline as the rest.
LLM-assisted test generation, intelligent test selection, self-healing automation and predictive quality analytics, reducing manual test design effort by approximately 40%.
NarraTwin AI delivers grounded scripts with citations, claim evaluation, consent checks and a release gate that runs before generation, not after.
- services & APIs
- trust boundaries
- AI workflows
Evaluate — coverage that means something
Coverage counted against acceptance criteria, not files. For AI systems the same discipline applies to faithfulness, relevancy, hallucination and retrieval precision — measured, baselined, and defended.
Automation coverage increased from 65% to 95%, cutting regression execution effort by approximately 25% and flaky tests by approximately 20%.
EvalAxis scores five metrics — faithfulness, answer relevancy, hallucination, context precision, context recall — against a committed baseline. Its default judge is deterministic and runs offline: it proves the mechanism, not semantic quality.
- faithfulness
- answer relevancy
- hallucination
- context precision
- context recall
Release — governance with a spine
A release should be blockable for a stated reason and releasable on stated evidence. CI/CD and infrastructure are the plumbing; the gate policy is the judgement.
Targeted release workflows reduced cycle time by approximately 25%.
CiteVyn promotes a new retrieval index only when the newest completed evaluation run clears a stated pass-rate threshold — one transaction, gated, with an audit event on both the clean and the forced path.
- pass — within threshold
- block — regression
- warn — drift detected
Operate — degraded modes are a feature
Telemetry, canaries, alarms and runbooks decide how quickly a bad day ends. In AI systems the same instinct becomes visible mode labelling: the user should be able to tell whether what they are seeing is current, held, or absent — never a fallback disguised as the real thing.
MTTD reduced by approximately 35% through telemetry and dashboards; production incidents reduced by approximately 20%.
SaafSaans labels every reading live, held (shown as CACHED — a real earlier measurement re-served because the upstream failed, not stale data), or none (shown as NO READING, never disguised as a fallback). Quorum-AI discloses a degraded run — forced into local simulation rather than spending against an untrustworthy cost ledger — and estimates cost before a run starts, not after.
- live
- held / cached
- no reading
- degraded
Learn — production teaches the requirement
The loop only closes when what happened in production edits what gets specified next: known limitations written down, baselines kept honest, evidence deciding what a system is allowed to claim — not enthusiasm.
Defined enterprise-wide quality engineering standards and governance while mentoring 11+ quality engineers and SDETs globally.
project-doc-skills — eight independent skills that turn a project into documentation, each in its own mode, with an independent review gate before anything publishes.
- known limitations
- not-claimed list
- next evidence needed
All eight stages are in the page source and readable without JavaScript or animation.
Evals and observability, specifically.
A calibration harness and a drift test that both admit their own limits.
Judge agreement scored with Cohen's κ, PABAK, and Spearman correlation — the labels it checks against are explicitly not blind human labels, its own decision record says so. Separately, a Kolmogorov–Smirnov test watches live output against a stored baseline, with an alert on a floor breach — demo-proven, not yet wired to live traffic.
Tracing with the secrets stripped, and two ways of separating a blocking gate from an advisory one.
OpenTelemetry and Prometheus across two systems, secrets removed before it leaves the process. One repo's judged eval only runs on a pull request labelled full-eval, since it costs real money; another enforces a fixture-backed eval suite as one of eleven contexts checked live against the branch-protection API.
Published, not just practiced.
One of the skills above is a public repository — the instructions themselves, not a description of them.
- project-doc-skills — architecture-and-decisions and doc-critic, the two skills behind this site's own documentation practice.
- .github — the contribution and security templates every other repository here inherits from, not a skill.