Skip to main content

Selected Work

The systems I build and operate: healthcare integration tooling, Kubernetes platform engineering, and the agent delivery pipeline that builds both under gates I can audit. Everything public links to source.

fi-fhir: healthcare integration engine

  • Go
  • open source

Format-agnostic pipeline that parses legacy healthcare formats (HL7v2, X12 EDI, CSV) into semantic events like patient_admit and lab_result, then routes them through configurable workflows. Integration logic reads in workflow terms instead of PID.3.1.

Includes staged parsing with Source Profiles, FHIR R4 output, a GET /metadata CapabilityStatement checked against implemented mappings, and CI coverage gates. The Mapping Studio and remaining 1.0 release work are documented in the case study.

The playground runs the real fi-fhir parsing and FHIR mapping kernel, compiled to WebAssembly, in your browser: paste an HL7v2 message and get its segments, semantic events, diagnostics, and FHIR transaction Bundle. It ships synthetic samples only, and nothing you paste leaves the page.

Live demo: the fi-fhir IDE with a preview-only identity — paste synthetic HL7v2, no data is stored.

edilint: pre-send linter for EDI and healthcare flat files

  • Go
  • open source
  • Apache-2.0
  • v0.6.0

Single static binary that gates outbound files in CI or a send script: X12 envelope integrity with control-number and segment-count recounts, invisible and lookalike character detection, terminator consistency, declared-versus-actual trailer counts, and fixed-width layout drift.

48 rules across X12, HL7v2, EDIFACT, delimited, and fixed-width files, with every X12 rule cross-referenced to the TA1 or 999 code a trading partner would return for it. v0.3.0 added fmt, fix, diff, and stats subcommands, plus an MCP server so a coding agent can lint the files it generates. Zero dependencies, JSON and SARIF output, CI-friendly exit codes, tested on a three-OS matrix. Built with agentic engineering under adversarial review, and the review is in the commit history.

FlexInfer: Kubernetes-native GPU scheduling and model serving

  • Go
  • open source

A Model resource becomes a serving lane: FlexInfer picks a GPU, stages the weights, starts a backend (vLLM, llama.cpp, diffusers, and others), and puts the lane behind one OpenAI-compatible endpoint that scales it to zero when idle. Lanes that share a card take turns by priority and demand, and LoRA adapters hot-load into vLLM lanes.

It runs my cluster's LLM, embedding, rerank, TTS, image, and video lanes on two AMD RX 7900 XTX (gfx1100) cards, a Radeon VII (gfx906), and a GTX 980 Ti. On gfx1100 the Qwen3.8-27B workhorse moved to vLLM 0.32 with MTP speculative decoding and RDNA3 segmented attention, and decoded at 68.5 tokens/s at 2K context when it was promoted on 2026-10-26, up from about 40. On gfx906 the Radeon VII serves retrieval and speech as llama.cpp lanes in one runtime pod and swaps models in 3.5 to 9.2 seconds. Every runtime change goes through a canary, a soak, and quality gates before it takes traffic.

Loom and Mills: MCP ecosystem and an agent delivery pipeline

  • Go and TypeScript
  • open source

A Go daemon plus VS Code and Zed extensions that manage Model Context Protocol server lifecycles, sync configuration across supported AI coding tools, and provide persistent agent memory across sessions. This is the tooling I use to build everything else on this page.

Mills sits on top: a Kubernetes operator that plans backlog in a council, implements in sandboxed pods, runs gates (scope, protected paths, secret scan, tests pinned to the pushed commit), opens the merge request, watches CI, and merges under a policy held in GitOps with a kill switch and per-repo budgets. Every run leaves stage records with cost, gate verdicts, and the commit that was actually tested. fi-fhir and edilint are now maintained through it.

In the 30 days to 2026-10-30, Mills ran 913 times, merged 417 changes, and sent 290 runs to me tagged by cause (transient failures, code, configuration, infrastructure, and a few others). The factory page shows the sanitized record.

Integration Command Center: workbench for integration programs

  • Python

Self-hosted system of record for vendor-integration work: projects and vendors, artifact capture with SHA-256 evidence hashing and data-classification levels, decisions, risks, milestones, Gantt/Kanban/status views, hybrid FTS5 plus semantic search, an MCP server surface for AI agents, Prometheus metrics, and a formal threat model gating what data may enter. I used it daily to run real multi-vendor integration programs.

Code private (operational data). Architecture walkthrough and synthetic-data demo on request.

daemon: personal data platform

  • Python/FastAPI and React/TypeScript

Roughly 300 routes handling multi-source ingestion (Gmail classification pipeline, feeds, calendars), five storage tiers (Postgres, Neo4j graph, Qdrant vectors, Redis, MinIO), LangGraph agent workflows for document composition, and ARQ background jobs. About 140K lines.

It runs my job search: lead intake from job boards and mail, a resume registry that composes tailored variants, a claims linter that flags any line in a letter or deck stated more strongly than its source, and an interview prep queue.

Private (personal data). Happy to walk the architecture.

The platform underneath

Everything above ships through a homelab I operate as production: K3s on Harvester, Flux CD GitOps, Longhorn storage, Harbor registry, self-hosted GitLab with CI runners, and a Prometheus/Grafana/Loki observability stack.

Until October 2026 I led the Payment Integrity and Payments integrations team at a Medicare Advantage payer. Those systems belong to that employer; the details are mine to tell in interviews. The code above is what you can inspect today.

Selected Work | Cody Blevins