Matthew Francis.

Matthew Francis · Grand Junction, Colorado

I build systems that hold up under scrutiny.

Full-stack engineer in regulated health-tech. I lead engineering on a HITRUST-certified 340B platform, run a distributed team across time zones, and build verification-first agent tooling with a human gate on everything that matters.

matthew@francis.best GitHub LinkedIn
6+years shipping production systems
3years leading a HITRUST-certified 340B platform
11engineers led across the platform's life
99.9%measured uptime across the platform's last 12 months
01

Work experience

2019 · Present
01 / Technical Lead Jul 2022 · Present

SaaS Factory · Bespoke Creations · my employer

The consultancy I have worked for since 2022, and the firm through which the engagements below are delivered. I am its lead engineer for demanding client work: accountable for architecture, code quality, release, and production reliability, and the technical point of contact for clients.

Engagement · Jun 2023 · Present

AlineaOne (Alinea Group) · Technical Lead

Multi-tenant 340B drug-pricing and compliance platform under HITRUST and PHI constraints.

  • Lead the platform's engineering and QA team, currently seven engineers across time zones and eleven over the life of the platform, owning code review and delivery from ticket to production
  • Own end-to-end technical direction for a schema-per-tenant Django and React platform on AWS EKS: architecture, the full release pipeline, and production reliability
  • The platform held 99.9% measured availability over the trailing twelve months, with each of its eight flagged hours traced to a known deploy or probe event.
  • Building the platform since 2023, at 1,000+ commits a year across its codebase and infrastructure. Hands-on from EDI ingestion and multi-tenant data isolation to the release pipeline and HITRUST evidencing.
  • Built and hardened the EDI ingestion pipeline (832/810 pricing and invoice files over SFTP) with per-tenant routing and idempotency, feeding regulated compliance reporting
  • All infrastructure defined as code in Terraform, with a dev to staging to prod CI/CD pipeline
  • Translate ambiguous regulatory and client requirements into features that stand up to audit
  • One of four core leaders who guided the platform through its i1 HITRUST certification — owning technical readiness and implementing and evidencing the controls that carried it through assessment
  • Named the on-call Technical Lead for disaster recovery in the platform's HITRUST-attested policies and procedures.
DjangoReactPostgreSQLAWS EKSTerraformHITRUSTEDI 832/810

Engagement · Jun 2025 · Present

Suncoast Venture Studio · Infrastructure Lead

Infrastructure lead for a multi-venture SaaS portfolio, operated through infrastructure-as-code.

  • Own and operate the shared Kubernetes (EKS) platform running the portfolio, plus all Terraform IaC
  • Manage release pipelines, DNS, and cloud security posture: WAF rollout, Security Hub remediation, secrets migration to AWS Secrets Manager
AWSEKSTerraformGitHub ActionsSecurity HubWAF

Engagement · Sep 2022 · Sep 2024

Intuition AI · IT Admin

  • Trusted with administrator access to critical systems
  • Established an organization-wide password manager, improving security
  • Implemented automated CI/CD to speed development and stabilize production applications
02 / Automation Engineer Dec 2019 · Mar 2022

Independent · Automation Engineer

Freelance automation engineer for business clients, built web scrapers, data-extraction and automation scripts, and operational tooling that replaced manual, repetitive workflows. Let the robots do the robot work, so the humans can do the human work.

02

Projects

Personal builds · not client work

Project 01 · Multi-agent engineering system

A multi-agent engineering system, human-gated by design.

A personal build I designed and operate: a hierarchy of long-running LLM agents, an orchestrator routing to domain directors routing to workers. It ingests real-time signals, dispatches scoped work, and carries delivery cycles under guardrails and human gates. The leverage is real, and so is the human in the loop.

  • Real-time signal ingestion via webhooks, pollers, and schedulers across email, chat, CI, and ticketing
  • Tool operation over an MCP layer with a central dispatcher and per-domain routing
  • Persistent episodic memory, distilled from every working session and recalled on demand
  • Guardrail hooks with hard gates on compliance-sensitive and irreversible operations
  • Verification layers that treat model output as a hypothesis to confirm, never as an answer
  • Measured, not assumed: a first audit of 67 logged agent claims found 61 resolvable against a primary source. Nine of those (15%) were wrong on a load-bearing point. Verification caught eight of the nine before anything shipped. The one known escape is why the human gate stays. Measurement is ongoing as the logged corpus grows.

Inbound signals

WebhooksPollersSchedulersChat
Orchestrator
Draft PRsReview notesDiagnosesVerified checks

A delivery cycle, gated where it counts

Diagnose Implement Test Review Merge Deploy Verify

In daily use on my own work: triaging tickets, drafting and reviewing code, watching CI, and moving infrastructure changes through plan, review, and staged CI/CD. Model output is treated as a hypothesis, verified against the primary source before anything ships, with hard gates on compliance-sensitive operations.

One recent cycle, end to end: a failing CI run tripped a webhook after midnight. The orchestrator routed it to a domain director, which spun up a worker in an isolated worktree; the worker reproduced the failure, traced it, wrote a fix, and opened a draft PR unattended. A separate review agent checked the diff against the ticket's acceptance criteria and caught that the worker's confident root cause was wrong — the real regression only surfaced against year-old data, which showed up only because the verification step pulled the actual records instead of trusting the diagnosis. It was corrected before reaching me; I approved one merge in the morning and staged CI/CD shipped it. The system did the diagnosis, reproduction, and first draft overnight; the verification layer killed a plausible-but-wrong hypothesis; the human gate was a single approval, not a morning of triage.

Project 02 / Founder Live at cashcraft.org

Cashcraft · cashflow-first personal finance

A personal budgeting web app I design, build, and run myself: product design, data model, and the allocation and projection engine, end to end.

  • React, Vite, and TypeScript frontend on a Supabase (PostgreSQL, auth) backend
  • Live bank connectivity via Plaid, continuous deployment on Vercel
ReactTypeScriptSupabasePlaidVercel
Project 03 / Open source 63 stars · 9 forks

vector.lua · 2D/3D vector math for Lua

A small vector math library for Lua, with a PVector-style API modeled on Processing — the add, scale, normalize, and dot/cross product operations that games and creative-coding sketches reach for constantly.

LuaOpen Source
03

Stack

Daily drivers

Languages

  • Python
  • TypeScript / JavaScript
  • SQL
  • Bash

AI & Agents

  • Multi-agent orchestration
  • MCP servers & tooling
  • Frontier model APIs
  • Episodic memory & retrieval

Cloud & Infra

  • AWS (EKS, RDS, S3, WAF)
  • Terraform
  • Kubernetes
  • CI/CD pipelines

Product

  • Django & REST APIs
  • React + Vite
  • PostgreSQL
  • Multi-tenant architecture
04

How I work

A

Model output is a hypothesis, never an answer.

Every agent claim gets verified against the primary source before it ships. Trust is earned per claim, with receipts. That discipline is built into the tooling itself.

B

Own the architecture and the implementation.

I use coding agents where they create leverage and step in directly where precision and judgment matter. Ambiguous 0-to-1 problems become reliable production systems.

C

Compliance is a design input.

Three years of HITRUST, PHI, and 340B work: idempotent ingestion pipelines, staged releases through reviewed CI/CD, and access scoped to the least each service needs.

Available for the right problem

Have a hard problem in regulated software?

github.com/automattf linkedin.com/in/automattf