What I build
Systems I've built and run.
Flagship
Tangerine-Home
A self-hosted AI platform for a whole household — a WhatsApp butler, shared documents, daily reminders, and a private way to pass on what matters when it's needed. Same data whether you open the app or message the butler. Running in daily use since 2025.
Enter for more videos →Platform & infrastructure
The layer everything else runs on — built, operated and monitored by one person.
~90s unattended recovery
Self-healing deployment pipeline
Automated rollback on failed deploys, with paired sensors and actuators for each recurring failure mode. Anything needing judgement escalates to an agent that drafts the fix but cannot ship it.
8+ services · one edge
API gateway & authentication layer
One edge fronting the platform. An opaque per-member key is exchanged for a server-held credential and brokered into per-user tokens that never leave the host.
5 isolated tenants · one host
Multi-tenant provisioning
Scripted provision and teardown of fully isolated environments — dedicated databases, deterministic ports, namespaced identity for shared services. A tenant you can’t remove cleanly was never isolated.
94 containers · verified nightly backups
Self-hosted operations
Zero-trust tunnel ingress instead of open ports, automated TLS, and backup verification that derives what it expects from the system itself — so a new service can’t go silently uncovered.
Policy by config, not code
Governed remote access for agents
An installable agent giving an assistant real hands on machines beyond the server. Capability set by configuration, command guards proven by test, and a data-residency model built for client work.
AI & agents
Agents with real tools and real limits — built from first principles, not assembled from a framework.
~50% inference cost cut
Model routing & cost control
Task-based routing across cloud providers and a local tier — deterministic work runs locally, judgement buys a frontier model. Services name a role, never a vendor.
Open source
Agent-to-agent conferencing
Distinct AI agents on different surfaces holding a recorded conversation — one shared transcript format over per-pair transports.
View the code →Products & delivery
What the platform looks like to the people using it — and work built for someone else.
Client engagement
Demand & supply planning
Built for a 30-year-old precision manufacturer reconciling demand against supply by hand across dozens of printed pages. Netting computed event by event rather than by month, exceptions first — exposing three structural errors in the incumbent process. Designed and built; pending client validation.
8 productions
Deterministic demo pipeline
Every clip here is a real, redacted rebuild of the actual interface — self-playing HTML captured headlessly, then scored. No generative video, so nothing on screen shows a product that doesn’t exist.
Architectural framework
Four maps of how it is put together — each drawn in the industry’s own vocabulary, and each ending on what the standard tool would add.
How code ships
How code reaches the households
Checks before the push, tests, rollout and verification — and an automatic rollback when it goes red. A person reviews every fix before it ships.
See the map →How the agent works
How the agent runner works
An autonomous agent that works its own queue: a resumable state machine with a hard human approval gate, mapped onto agent-framework vocabulary.
See the map →How services run
My platform mapped onto Kubernetes
Each hand-built piece placed against the Kubernetes primitive it matches — and where it doesn’t. Built from need; I don’t run Kubernetes.
See the map →How you know
How you know the agent did it
Proof is made at the tool, checked at the boundary and audited in the reply — so “I don’t know” never looks like “it worked”.
See the map →Building something similar, or think I could help? I'm always up for a conversation.
Email me →