Six weeks to release · two prior attempts had failed
The front end behind Elavon’s Fusebox gateway
Elavon’s front-end payment controller had been attempted twice without
success. Brought in to rescue it; the third build held. It was a single-process
event engine in C: one poll() reactor over non-blocking sockets,
with socket readiness, timer expiry, and OS signals all reduced to a single
currency, an event driving a per-connection state machine. Around it, a plugin
architecture carrying fifteen point-of-sale protocols, and a pre-forked worker
pool that passed live connections across process boundaries by file descriptor.
First cut in three weeks, first release in six. Later rescued and finished the
batch settlement engine as well: C over embedded Sybase SQL, settling across a
dozen card networks, plus the 64-bit Linux migration.
Scope: this was a large team
effort over several years. The front-end controller and the settlement engine
were the components I owned; the authorization switch itself was other
people’s work. What shipped is the
Fusebox
gateway, which Elavon still sells — now advertising “certify
once” access to seventeen third-party processors, against the fifteen
protocol plugins it launched with.
Ten minutes · after three weeks on site
A stalled call-center cutover, diagnosed in ten minutes
A cellular carrier’s contact-center project was dead in the water on a
Rockwell CTI switch. The switch vendor had flown engineers in and kept them on
site for three weeks with no fix. Put the software under a debugger, found a
large local array overflowing the process stack, raised the stack allocation in
the bind command, and re-linked. The system ran.
Live system · seven latent defects surfaced
New development on a running claims switch
A real-time pharmacy-claims switch on HPE NonStop was being migrated to the
cloud and still had to grow while it moved — adjudication happens while the
patient waits at the counter. Designed and delivered a vendor-authentication
feature end to end: each claim authenticated by its originating vendor and routed
on that result. Getting it working surfaced seven separate defects scattered
across the line configuration, the comms-layer JSON parser, and the inbound TCP
handler.
Shipped to production · February 2026
Mainsail — an AI platform that operates the business, not just describes it
Built and delivered in February 2026:
a multi-tenant Python platform of roughly 110,000 lines, architected as a set of
self-describing tools exposed to a language model over the Model Context
Protocol, so the model operates the company’s systems instead of
describing them. Underneath sit deep Salesforce and Accounting Seed integrations
with a local SQL mirror for fast analysis.
It delivers contract-level cash-flow forecasting
out to the KC-46 tanker program’s 2027 horizon, a rolling P&L with
anomaly detection, reconciliation between the CRM and the ledger, and an
exploratory-analysis agent that runs its own tool-use loop against open-ended
questions. Built concurrently with a live NonStop migration engagement. Platform
IP is held by Mainsail Labs — product
detail at mainsail.is and
zalitylabs.ai.
