RaffleDock
A raffle platform for schools and nonprofits: card payments, printed QR tickets, per-student attribution and live sales reporting.
raffledock.com →Scott Morris
Private AI running on hardware you own. Field apps that work without a signal. Data systems where the guardrails are built into the plumbing, not the policy document.
Private by design
The language models run on a GPU in your own rack — not someone else's API. Indexing, retrieval and answers all stay inside your network, and it keeps working when the internet doesn't.
Structural guardrails
An assistant on a read-only account physically cannot write. Compensation fields stripped server-side never reach the device that would render them. Guardrails belong in the data layer, where no prompt and no clever user can talk them out of it.
Agentic engineering
Eleven specialised agents in production with persistent memory, a shared task board and automated weekly reviews. Not a demo — it runs every day, on local hardware.
Regulated data
Payroll, medical and personnel data, handled with least-privilege access, data minimisation and tamper-evident audit trails. Designed to HIPAA constraints — practices, not a certificate on a wall.
Mobile
Native iOS for crews outdoors, wearing gloves, often with no signal. One binary whose modules the server decides at launch — granting someone a new tool is a permission change, not a release.
Payments
A raffle platform taking card payments for schools. Every purchase path is tested before it goes near a parent's card, because here the failure mode isn't a bug report — it's someone's money.
How it's secured
Retrieval and generation on owned hardware. No third-party model calls, no documents leaving the network, no API key to leak.
Analysis runs against read-only mirrors. Write access is a deliberate, separately authorised exception — never ambient, and never blanket-approved.
Sensitive columns are withheld server-side, so the client cannot render what it was never sent. Enforced in the query, not requested in a prompt.
Answers cite the record they came from, or the assistant says it doesn't know. No confident invention over a company's own data.
Credentials read from the environment, never written to disk, never committed, never echoed into a log or a working file.
Private-mesh networking instead of public endpoints. If a service has no reachable front door, it has no front door to attack.
Selected work
A raffle platform for schools and nonprofits: card payments, printed QR tickets, per-student attribution and live sales reporting.
raffledock.com →Built on it and running: printed QR tickets, card checkout, automatic receipts, and per-student sales the office can actually reconcile.
mues-raffle.com →Design and build of the public site for a clinical-AI company.
quietsignaltech.com →Retrieval over a company's own documents and databases, answered by a model on its own GPU. Nothing indexed, sent or stored off-site.
Native iOS tools for construction crews — supply walks, ordering, daily logistics. Built for gloves and bad signal.
Field-captured hours flowing into a production accounting system, read-only by default, with every money-affecting call reviewed before it runs.
On device
I take on select side projects. If you've got a problem that needs real software rather than another spreadsheet, say hello.
Get in touch