Matthew BobackToronto, Canada
I build backend systems and the interfaces that make them possible to operate.
Currently Platform Engineer at MnB Security, building an offline-first access-control and video platform and an evidence-checked research agent. Before that, two products of my own that are still running.
The terminal beside thisbelow runs a live StageFlow scan — paste any public URL to try it.
$ enter any public URL and press Enter to run a live scan
Physical
Doors that fail safe
An access-control platform where taps are authorized on site and the relay times its own relock. Validated on a 4-door bench and a 15-camera simulated site.
Agents
LLM output you can check
A research agent whose every quote is matched against the page it cites, on a Go runtime shared by three products.
Shipped
Products you can use today
StageFlow is open source with a live demo; AlchemizeCV runs in production. Both are inspectable, not just described.
What changed in the last few months.
- 2026-09Research agent: evidence-gated briefsA sandboxed LLM agent reads a company's public site and returns a cited brief. Every quote is checked against the captured page, a second model scores each citation, and a human makes the call.
- 2026-09Agent runtime and browser CLI extractedThe Go agent harness and the patchright-cli browser daemon moved out of AlchemizeCV into their own repos, now shared by three products under a one-way dependency rule.
- 2026-09StageFlow: scanner noise cut at the sourceScanned real sites, then fixed the scanners. This site went from 84 findings to 10; a 22-page scan from 575 to 218. Page discovery from sitemaps and crawls landed with it.
- 2026-09Managed Video on the access platformGo site agent, a relay that only copies bytes, single-use edge grants. A simulated site of 15 cameras kept recording through a 90-second cloud outage.
- 2026-08AlchemizeCV: 22 containers down to 8Removed Temporal, the dedicated workers and the streaming layer. Agent sessions now run on idempotency keys, sequenced transcripts and cursor polling. Architecture tests keep the old pieces out.
- 2026-07Gate A: device identity and a conformance gateHPKE-sealed credential bundles, three least-privilege service identities, and a Go tool that checks deployed broker and storage permissions against a hashed topology manifest in CI.
- 2026-07StageFlow: a stuck job, fixed properlyA lost pod-create response left a job pending forever. The orchestrator now adopts pods Podman already created, records exhausted deliveries, and sweeps jobs stuck in setup.
- 2026-06Access control on commodity hardwareOSDP readers and a Modbus relay module replace the vendor controller. Taps are authorized against local SQLite; the relay times its own relock, so software can't leave a door open.
Four systems, each written up for a skeptical read.
Two built for my current role at MnB Security and two products of my own. They show the mix I care about: backend architecture, operator-facing product work, agents with guarantees in code, and deployment decisions that survive real use.
MnB Access & Video Platform
A managed access-control and video platform for commercial buildings, built solo at MnB Security. Each building runs a Go site computer that authorizes card taps against local SQLite and pulses commodity relays through OSDP and Modbus, so doors keep working through internet outages. The cloud (React Router portal, Hono gateway, Postgres, NATS JetStream, MinIO) manages sites, people, cards, doors, and Managed Video, and rebuilds its view from what each site reports.
0
Inbound ports at the site
≤ 1
Door open per request
4-door
Validation bench
MnB Research Agent
An evidence-checked research assistant for a commercial-security business. A sandboxed LLM agent reads the public websites of firms that might subcontract installation work and drafts a cited brief; a final-answer gate rejects any quote that is not verbatim in a captured page, an independent verifier model re-reads every cited passage, and a person makes every decision. Built as three private repos: a Go agent runtime, a TypeScript browser tool, and the product on SQLite.
10
Model Providers
582
Harness Tests
0
Quotes Accepted Unchecked
StageFlow
StageFlow is an open-source, self-hostable frontend quality platform. It runs seven scanners — accessibility, performance, SEO, links, security headers, social metadata and content quality — as one pipeline behind a single report contract, and remembers a baseline per project so every scan can answer the question that matters in CI: did this change make the frontend worse?
7
Scanners
MIT
Open source license
3
Clients: web, CLI, API
AlchemizeCV
AlchemizeCV is a live job-application product. A user builds a profile, an AI coach turns evidence into bullets, a reviewed catalog of 1,024 job boards and a research chat find roles, and the app prepares, renders and submits Greenhouse applications only after the user approves each one. In August 2026 I replaced Temporal, dedicated workers and browser streaming with five long-lived services and a private agent harness, taking production from 22 containers to 8.
8
Containers, Down From 22
1,024
Job Boards
1
Submit Per Approval
What I tend to own when the work matters.
I am strongest where architecture and delivery meet: clear system boundaries, understandable UI, and enough operational rigor that the work does not collapse under real usage.
Backend systems that stay inspectable
I like event flows, explicit state boundaries, and operational surfaces that explain what the system is doing without tribal knowledge.
Agents whose output can be checked
When a model is in the loop I put the guarantees in code: sandboxes, typed contracts, evidence gates, and a human decision at the end.
Product interfaces for technical work
The UI work matters because operators need trust. I care about hierarchy, evidence, and interaction clarity as much as endpoint wiring.
Testing that supports shipping
Simulated hardware, browser automation, accessibility checks, and integration coverage are part of delivery, not cleanup after the fact.
Single-host deployment discipline
I work comfortably close to the machine: containers, reverse proxying, systemd, WireGuard, and release workflows that make rollback and debugging sane.
$ hire --role backend,platform,full-stack
Looking for someone who can own both the system and the proof around it?
A strong fit for roles where the work includes architecture, operator UX, testing discipline, and deployment ownership — instead of splitting those concerns apart.