Shipped, and checkable.

Every client here is described by sector rather than named. Two are shipped products in daily use, two are enterprise delivery under NDA, and one is a hackathon build that says so. A name goes back on a study when that client has agreed to it in writing, and not before. What is described in every case is the situation, the constraint that shaped the architecture, and the part I owned.

Small business · Web platform migration

Shipped, client not named

A family flooring contractor off WordPress and onto something it owns

Situation
A family flooring contractor in New York City had a WordPress site that was slow, hard to change, and held together by plugins. The gallery lived in the WordPress media library, the quote form went through a third party, and nobody on the business side could tell whether the site produced leads.
Constraint
Years of accumulated search history could not be thrown away. Every address the old site had ever answered needed to keep working, and the business needed to be able to go back if the launch went wrong.
What I built
A statically built Astro and Tailwind site, deployed to Hostinger over FTPS by GitHub Actions on every merge. 297 legacy URLs mapped to single 301 redirects, smoke-tested across 1,440 requests before and after launch. The gallery migrated out of WordPress into typed local data with real alt text. Google Business reviews pulled at build time. GA4 wired so a submitted quote registers as a lead. Build-time gates that fail the build on a broken link, a wrong canonical, a malformed sitemap, an uncovered redirect, or two service pages that have drifted into near-duplicates.
What I owned
All of it: the audit, the architecture, the build, the migration, the deployment pipeline, the launch run sheet, and the rollback plan.
Result
Live at 12:41 PM New York time on 26 October 2026, with a 30-day rollback window held open to 25 November. The old WordPress install was archived rather than deleted, out of the web's reach. The first release after launch shipped the same afternoon.

297 legacy URLs preserved, 1,440 redirect assertions green, 32 of 32 CDN edge checks serving the new site

  • Astro
  • GitHub Actions
  • Migration
  • Technical SEO
  • GA4

Read the full case study

Education technology · Collaboration

Shipped, client not named

An adaptive learning engine that has to be right about a child

Situation
A small education-technology team is building an adaptive learning app for homeschooling families: a child does 15 to 30 focused minutes of math and language a day, and the system adjusts. The hard part is not the app. It is an engine that scores every answer honestly, decides what that implies about a concept, and explains itself to a parent who did not ask for a dashboard.
Constraint
Everything the engine asserts gets read by a parent about their own child, so a grading bug is not a rendering bug. Getting a question wrong in the child's favour is as bad as the reverse, and a memo that reports the wrong count destroys the trust the whole product runs on.
What I built
Work across the Next.js application and the curriculum engine. Per-concept difficulty adaptation. Keyword grading corrected to whole-word matching, which killed the substring false positives. Figure generation that fails closed instead of rendering a zero-length highlight. Parent memos that report the real count. A year-plan page tested against the actual curriculum spine, not a fixture. The content workflow runs on LangGraph with deterministic quality gates, LLM-assisted review, a human-review interrupt, per-run cost ceilings, staging isolation, and post-deployment smoke tests.
What I owned
Engine and content-engine work inside the team, delivered as small reviewed changes: one task file, one branch, one pull request, each with the check that proves it.
Result
A working product under active development, built by a small collaboration rather than a single team. Three parts: the learning platform, a community education program putting it in front of real families, and the knowledge layer a student keeps when they leave.

Metric pending NDA review. The figure exists; clearance to publish it does not yet.

  • Next.js
  • Adaptive learning
  • Content engine
  • Evaluation

Read the full case study

Global logistics & enterprise IT · AIOps

Enterprise delivery, under NDA

Agentic operations platform for SAP and ServiceNow

Situation
Alert volume across SAP and ServiceNow was outpacing the operations team. Triage was manual, inconsistent between shifts, and the institutional knowledge for handling an incident lived in whoever had seen it before.
Constraint
Global operations, so nothing could be a black box. Every automated action needed to be traceable after the fact, and a human had to stay in the loop on anything consequential.
What I built
A Microsoft Agent Framework platform handling SAP alert processing, ServiceNow workflow automation, AI-assisted incident triage, and RAG-backed recommendation workflows, instrumented end to end with OpenTelemetry. The move off the earlier LangChain implementation was done against parity tests. The new path had to produce the same recommendations as the one it replaced before anything was cut over.
What I owned
Agentic orchestration through the full development cycle: design, architecture, build, and deployment. The recommendation engine at the centre of it: resolution, priority, and assignment-group recommendations, with confidence thresholds, fallback behaviour, and an explainable output.
Result
Deployed and running across global operations.

Metric pending NDA review. The figure exists; clearance to publish it does not yet.

  • Microsoft Agent Framework
  • OpenTelemetry
  • RAG
  • SAP
  • ServiceNow

Read the full case study

Enterprise IT · GenAI service management

Enterprise delivery, under NDA

The assistant that became the platform

Situation
Incident management, SLA tracking, and notification handling were spread across tools and people, with no single surface an engineer could ask a question of.
Constraint
It had to prove its value on real ITSM workflows before anyone would fund a platform build on top of it.
What I built
A LangChain work assistant covering incident management, ServiceNow workflows, SLA tracking, and autonomous notification handling.
What I owned
The GenAI foundation, end to end. It seeded the Azure-native platform above, which now serves global enterprise operations.
Result
Earned the mandate for the larger platform.

Metric pending NDA review. The figure exists; clearance to publish it does not yet.

  • LangChain
  • ITSM
  • ServiceNow
  • SLA

Read the full case study

Hackathon build · Infrastructure automation

Hackathon build, not production

Natural language to reviewed, deployable infrastructure

Situation
Infrastructure requests queue behind the small number of engineers who can write Bicep and Terraform correctly. The bottleneck is not intent, it is translation and review. A hackathon was the chance to ask what a governed version of that would have to look like.
Constraint
Generated infrastructure code cannot be trusted on sight. Any design worth building assumes the model is wrong sometimes. Nothing reaches an environment without deterministic validation, a standards and security review, and an explicit human approval.
What I built
A multi-agent pipeline spanning consultation, code generation, deterministic validation, standards and security review, pull-request preparation, plan review, and a human deployment approval gate, across Azure AI Foundry, Microsoft Agent Framework, MCP servers and GitHub workflows.
What I owned
The architecture narrative and the shape of the pipeline: where the gates go, what each stage has to prove before the next one runs, and why engineer-controlled release is the point of the design.
Result
A hackathon entry, running since April 2026. It was built and demonstrated. It has never handled production requests, and it is on this page as architecture.

No production metric. This was built and demonstrated at a hackathon and never handled real requests.

  • Terraform
  • Bicep
  • MCP
  • Security review
  • Human-in-the-loop

Read the full case study

Numbers matter more than narrative. I would rather leave a gap than fill it with something I cannot source. Each gap above says why it is there, and it stays until a figure is cleared to publish.


Systems I've shipped.

Public platforms, private tools, collaborative builds, and field experiments. The status on each is the real one.

Open source · MIT

Zen Agent Skills

A portable library of agent skills, one harness-agnostic SKILL.md per skill, that installs into Claude Code, OpenCode, Cursor, and VS Code or Copilot from a single source. The rigor sits in two places: a written contract before the work and independent verification with evidence after it.

A canonical skill model with thin adapters per harness, Python tooling for installation, adapter generation and schema validation, and cross-platform CI on Linux, macOS and Windows. No skill ships cold. One is accepted after it has been used on real work and refined from what that use exposed. The catalog is smaller than it could be.

  • Python 3.11+
  • Claude Code
  • Cursor
  • OpenCode
  • No dependencies

Content infrastructure · Private monorepo

Zen Solutions Studio

The product studio behind the practice: one uv workspace holding every pipeline, agent, and tool, with shared infrastructure in packages and MCP tool servers in agents. Content OS runs the production pipeline from transcript to cut, with a human gate at each stage.

Recent work: a standalone YouTube media tool with playlist handling, ID3 tagging and clip extraction, a music-player runbook, and a live-stream guard.

  • uv workspace
  • MCP
  • WhisperX
  • FFmpeg
  • n8n
Active

Personal tool · Knowledge architecture

Savitri

A local-first second brain built for daily use, and used daily. Obsidian is the writing surface. Retrieval is hybrid over SQLite FTS5 and vectors with rank fusion, reachable by agents through a custom MCP server. Capture is automated. Filing something does not depend on remembering to.

Confidence-routed review and regression-driven repair, so the system says when it is unsure instead of answering anyway. Recent work: a page schema backfilled across 293 wiki pages, an Anthropic provider, and external source indexing.

  • SQLite FTS5
  • Hybrid retrieval
  • pgvector
  • MCP
  • Obsidian
Private preview

Community · Archery & bhakti yoga

Dhananjaya Dojo

The archery dojo in Alachua, Florida, serving children and adults pursuing precision and discipline. A session tracker with an AI sensei grounded in Bhagavad-gītā As It Is, Arjuna's own text, as the source of guidance.

  • Firebase
  • Session tracking
  • Gaudiya Vaishnava

AI platform · Production readiness

Ship It Right

A production-readiness and transformation advisory platform. Three assessment tracks, Builder, Business and PM, run parallel eight-domain analysis grounded in a session knowledge base, with a cited live-research report behind one click.

  • Firebase
  • Gemini
  • Assessment
Private beta

Community tool · Simulation & tactics

Pucker Point

A ballistic calculator and fire-mission management system for the ARMA Reforger realism community: call-for-fire workflows, range and deflection computation, and mission coordination for server operations. Systems thinking applied to a domain well outside AI.

  • Vite
  • Firebase
  • Ballistics
In development

Want this kind of work on your system?