Agentic systems, shipped to production.

An autonomous Jira-ticket-to-PR agent. It plans, implements in an isolated git worktree, verifies against per-repo tests / mypy / linters, pushes a branch, opens the PR, and comments back on the ticket — inference on AWS Bedrock.

• 5-stage workflow with human gates at Plan / Diff / Push — control without micromanagement.
• Hard per-run cost ceilings enforced in code: $1–25 budget, 100-turn cap.
• Shrinks a 2–4 hour ticket loop to ~10 minutes of review.

A Claude plugin that packages a ~254-service enterprise back-office codebase into 12 invokable skills — plan-ticket, find-service, implement, verify, prepare-pr… — backed by a ~280-file knowledge tree with per-service leaves.

• Skills cite source by file:line and refuse to invent facts when context is missing.
• Knowledge-graph retrieval so engineers drive ticket work end-to-end without re-reading the codebase each session.
• Turns tribal knowledge of a sprawling platform into a queryable, trustworthy assistant.

A production GenAI service live on Google Cloud that turns two companies’ PDFs into factually-grounded, layout-ready B2B marketing collateral — grounded generation with citations, deployed on Cloud Run with CI/CD.

• Every factual claim cites a source-tagged fact from the parsed briefs; the writer abstains rather than invents — verified by a claim-level faithfulness check.
• Word limits & layout enforced by a deterministic validate-and-repair loop in code, not by trusting the model; prompt-injection hard gate refuses poisoned PDFs, fail-closed.
• Cloud Run · Pub/Sub + DLQ · Firestore · GCS with CMEK · Terraform · Cloud Build CI/CD — 51 offline tests, measured cost ~$0.01 per article.

1,600 machine breakdowns, 20 sensors, only 40 expert labels. Proved the raw data has no cluster structure, found the one sensor carrying the failure signal, and propagated labels with honest, leak-free validation.

• EDA first: silhouette ≈ 0.04 at every k, HDBSCAN flags 100% noise — naïve clustering provably fails; ANOVA isolates the signal to a single sensor (F ≈ 25).
• Label Spreading on the selected subspace under nested cross-validation — reports the leak-free 0.50, and shows the leaky 0.60 as the trap it is.
• Calibrated confidence + an “unknown” bucket so novel failure modes route to experts instead of being forced into 3 classes; production design on Vertex AI Pipelines.

Precision energy intelligence for residential solar. A mobile-first PWA that turns raw inverter telemetry into insights a homeowner can act on — when to run the dishwasher, when to pre-charge the battery, which string is shading.

• Live solar → home ↔ battery ↔ grid energy-flow visualization with real wattages.
• 7-day production forecast from Open-Meteo shortwave-radiation modeling — no API key.
• SunSpec Modbus telemetry, battery-strategy modes, white-labeling, installable & offline.