Selected Work · 2024–2026
Agent orchestration architecture, MCP servers, Claude API integrations, and operational platforms. Each project is a working system, not a proof-of-concept deck.
Projects
Designed and deployed a full multi-agent operating system for a healthcare AI startup's operational intelligence layer. 20 agents across 5 domains — Intelligence Synthesis, Pipeline GTM, Data Operations, Quality Assurance, and Infrastructure — running in coordinated production with no external orchestration tooling.
Core architecture: a T1/T2 dual-tier instruction system that separates guard rails and evaluation routing (T1, loaded in agent settings) from full workflow procedures (T2, loaded on demand). T1 ensures consistent behavior and inter-agent handoff compliance at near-zero token cost. T2 provides full operational depth only when the task requires it.
Built a governance reconciliation matrix across all agents, formalizing every inter-agent handoff schema, escalation protocol, and human-gated control point. Signal routing pipeline delivers research intelligence from synthesis agents to decision surfaces in sales and leadership.
Three interlocking agentic infrastructure projects built from scratch. Together they form a working proof of the "You Are an API" thesis: extract intelligence locked in black-box systems and give users ownership of the layer.
Document intelligence system — a Claude-powered classifier. A 4-layer progressive cascade processes documents from filename only (L0) through PDF metadata (L1), document structure (L2), and targeted content envelope (L3). Privacy-first: L0 is the default, content reading requires explicit user permission. Validated against 10,836 real documents. Deployed on Fly.io with FastAPI, SQLite (WAL mode), and adapter-based ingestion from Google Drive and local filesystem.
Agentic identity provider — SSO for AI agents. A portable identity document (PID) that agents authenticate against instead of each building its own siloed model of the user. Scoped per-agent authorization controls what each agent knows and what it can do. Exposes a native MCP server interface so any MCP-compatible tool can request identity context. Built on WebAuthn passkeys with audit logging, scoped access, and revocation.
Portable AI memory server — normalizes conversation exports from Claude, ChatGPT, and Gemini into a unified corpus accessible as typed MCP resources. Deterministic SHA-256 IDs enable cross-platform deduplication.
A meta-analysis of AI prompting behavior across 18 Claude sessions (March 2026), scored across 8 dimensions: efficiency, delegation quality, pattern recognition, scope setting, instruction precision, context management, escalation judgment, and output specificity.
Built as a live interactive dashboard — rolling average timeline, radar chart, per-dimension breakdown — and converted into a Notion teaching module for the client team. The analysis itself demonstrates the pattern: extract the intelligence locked in operational behavior, surface it as structured infrastructure.
Score arc from 6.5 → 8.5 over 18 sessions documents measurable improvement in delegation quality and scope setting as the primary leverage points. Used to inform prompting standards across all active agent fleets.
End-to-end operations platform for a dance education company — built from concept to deployed in weeks, no framework, no build step. Six distinct modules covering the full student and instructor lifecycle.
Audio delivery is handled by a Cloudflare Worker backed by R2 object storage — protected routes, streaming delivery, zero server management. Onboarding flows use decision-tree logic to route new students to the right starting point. The choreography builder generates structured session plans from parameterized inputs.
Knowledge readiness system for a healthcare AI company's go-to-market team — 158 questions across 9 categories covering GTM, technical thesis, objection handling, competitor analysis, and agent system operations. Built to ensure every team member can speak the product fluently in any conversation.
Includes track filtering (Core / GTM / Ops), expert mode (tighter timing, fewer hints), a rotating knowledge drop that surfaces new content each session, and a score history system. Deployable in minutes, no backend required.
We design and build agent systems that operate in production — not decks.
Start the conversation →