00Shipping record
Jobs
delivered.
Three engagements, told plainly: what was broken, what we built,
what the numbers said afterward. Every figure below survived an invoice review.
The challenge
Meridian clears equity and futures trades for forty-odd brokerages.
Their matcher — a 2009-era Java estate — shed connections every time
volume spiked, which is to say every time it mattered. Two vendors
had already taken a run at it. Both proposed three-year migrations.
The exchange gave Meridian eighteen months before penalty clauses
kicked in.
The build
Four engineers, forty weeks. We rewrote the matching core in Rust —
lock-free order books, a custom binary protocol on the wire — with
Kafka carrying the event log and Postgres as the settlement system
of record. The old estate ran in parallel for twelve weeks behind a
shadow-traffic harness that replayed every production message into
the new core and diffed the results. Cutover happened on a Tuesday
at 02:00. Nobody outside the project noticed, which was the point.
The result
2.1 million messages a second sustained on commodity hardware,
p99 match latency of 180 µs, and fourteen months in
production — including two record-volume days — without losing a
message. Meridian's exchange penalties: zero. Their infrastructure
bill dropped by a third as a side effect.
- ClientMeridian Clearing
- SectorCapital markets
- Team4 seniors
- Duration40 weeks
- StackRust / Kafka / Postgres / k8s
2.1Mmsgs/sec sustained
180µsp99 match latency
0messages lost in 14 months
Lot No. W-02Cap 40THeavy Ind. SoftwareEst. 2019Inspected & passed
Case 02 — Logistics — 2024
Ferrostar
Logistics
Plate 02 — Switch points, night shift · routed traffic
The challenge
Ferrostar moves less-than-truckload freight across 118 terminals.
Load planning was a nightly batch job plus, in practice, seventeen
dispatchers with spreadsheets. The board committed — publicly — to
dynamic routing before peak season. That gave engineering fourteen
weeks, a date chosen in a boardroom with no engineers in it.
The build
Three engineers, embedded with Ferrostar's dispatch leads from day
two. A Go routing engine running constraint-based optimization over
31,000 daily routes, re-planning in under 40 seconds when a truck
breaks down or a dock backs up. Kafka streams terminal telemetry in;
Postgres holds the plan of record. We demoed working software every
Friday, and dispatchers' objections reshaped the scoring function
four times. That's not scope creep. That's the job.
The result
Shipped in week fourteen — the date printed in the board deck.
First full quarter: fuel spend down 11%, trailer utilization up 9
points, and the nightly batch job retired. The dispatchers kept
their spreadsheets for a month, then stopped opening them.
- ClientFerrostar Logistics
- SectorFreight / LTL
- Team3 seniors
- Duration14 weeks
- StackGo / Kafka / Postgres / k8s
14 wkspromised — zero slips
31kroutes planned daily
−11%fuel spend, first quarter
Lot No. W-03Cap 40THeavy Ind. SoftwareEst. 2019Inspected & passed
Case 03 — Insurance — 2025
Castellan
Insurance
Plate 03 — Forged block, machined true · made boring
The challenge
Castellan paid a systems integrator $6M over two years for a claims
platform that arrived late and half-finished: 34 production
incidents a month, a six-figure cloud bill for 40,000 daily claims,
and an internal team that had stopped believing the system could be
saved. The integrator's answer was a change order. Castellan's answer
was to call us.
The build
Rescue missions run in a fixed order: stabilize, then rebuild.
Weeks one and two were triage — circuit breakers on the flakiest
integrations, a real deploy pipeline, alerts wired to symptoms
instead of noise. Incidents dropped by half before we changed any
business logic. Then we carved the TypeScript monolith along its
actual fault lines: claims intake, adjudication, and payments became
three services with owned schemas, strangler-fig style, no big-bang
rewrite. Castellan's own engineers paired with ours the whole way —
the codebase had to outlive us.
The result
Nine weeks to stable: 34 incidents a month down to 2, p95 claim
intake from 11 seconds to 800 ms, infrastructure spend down
41%. Castellan's team runs the platform themselves now, with no
retainer required.
- ClientCastellan Insurance
- SectorP&C insurance
- Team4 seniors
- Duration9 weeks
- StackTypeScript / Node / Postgres / k8s
34→2incidents per month
800msp95 intake, from 11s
−41%infrastructure spend
