// selected work
Production AI, and the platform work that keeps it running.
A few builds I can talk about. Sanitized for public — customer names, internal product names, and specific figures removed; the patterns and outcomes are intact.
An AI operating system for enterprise pursuits
// outcome
- Enabled $0.5M+ and $1M+ enterprise engagements to close on platform-authored statements of work
- Compressed manual artifact production into command-driven workflows
- Productized an AI Well-Architected workshop; partner-funded engagement program added co-funding to a pursuit
// problem
Enterprise presales pursuits — discovery, qualification, proposals, SOW redlines, risk review, handoff — ran on manual effort and were inconsistent across a field team.
// what I built
- AI-native operating system on a single shared context store: 16 plugins, 95 slash commands, 22 skills, 6 subagents behind one operator surface
- Prompt-contract framework with versioned input/output schemas — violations surface in the dev console instead of degrading silently
- Evidence-citation guardrails: every claim carries a High/Medium/Low confidence rating and a dated source, or the output refuses to generate
- Automated test suite + plugin-validation and release automation; multi-provider routing (Claude primary, Bedrock fallback); prompt caching under a token budget
A pursuit microsite that closed Phase 1 in three weeks
// outcome
- Pursuit moved from ideation workshop to verbal Phase-1 commitment in three weeks
- Collapsed three serial deliverables (microsite, POC, workshop) into parallel motion
// problem
A complex enterprise AI-agent pursuit needed to move fast, and slide decks were not moving the client CEO or the hyperscaler co-sell partner.
// what I built
- Customer-facing microsite built in days on Claude-on-platform, anchoring every follow-up conversation
- AI-listener-agent architecture on Bedrock + ECS Fargate behind a WebSocket API Gateway, OpenSearch Serverless for RAG, Transcribe/Polly for voice
- A working POC built live in the ideation workshop (single-file Flask + Anthropic, ~3,700 lines, eight views) that pulled stakeholders into architecture faster than slides
Cutting a Claude Code corpus 19% with no loss of function
// outcome
- Took the corpus from 221K → 179K tokens (19% reduction) with no procedural steps or contract terms removed
// problem
A production multi-plugin Claude Code corpus had grown to 221K tokens — driving cost and context pressure every session.
// what I built
- A reproducible methodology measured with tiktoken / cl100k_base
- Seven patterns: rules dedup, trimming the right-skewed top decile of commands, removing repeated step counters, collapsing fallback boilerplate, one base template plus deltas, dropping redundant reload reminders — and leaving load-bearing specs untouched
// open source
$ git remote -v
Clean-room, tested, CI-green reference implementations behind the writing.
A bounded agent loop with three hard exits: iteration cap, cost ceiling, stall detection.
A provider-agnostic LLM client; swap backends with an env var, with a built-in smoke test.
Measure the token cost of MCP server tool schemas and emit a defer-loading manifest.
Measure a prompt/skill/command corpus, track reductions across snapshots, find duplicates.
Building something that has to survive contact with production?