Work

Every claim has a system
in production behind it.

Four engagements, written up with the numbers. Client names and references on request — most of our pipeline is their referrals.

01
Series-A fintech · 11 weeks

Figma file to SOC 2-ready beta

The problem

A funded team with a design system, a compliance deadline, and no platform. The previous contractor had left a prototype nobody could deploy twice the same way.

What we built

Next.js app on a typed API, Postgres with row-level security, audit logging wired from day one, and infrastructure in Terraform. Flagged releases from week three so compliance could review live rather than in screenshots.

11 wks
brief → beta
$3.1k/mo
all-in cloud spend
0
rewrites since launch
02
Logistics platform · 6 months

A monolith unpicked without a freeze

The problem

Eight years of Rails, a deploy that took four hours, and a team afraid to touch routing. A previous big-bang rewrite attempt had been abandoned after nine months.

What we built

Strangler-fig migration: the highest-traffic routes moved to typed services behind the existing edge, with dual-write and backfill for the data. Characterization tests first, so undocumented behaviour was pinned before anything moved.

4 hrs → 9 min
deploy time
31%
infra cost reduction
0 hrs
planned downtime
03
Healthcare AI · 14 weeks

Retrieval that clinicians actually trust

The problem

A model answering from a document corpus with no way to tell a good answer from a confident one. Pilot users had stopped using it after the second wrong citation.

What we built

Rebuilt retrieval with structure-aware chunking and re-ranking, every answer citation-linked to its source span, and an eval suite run on each change. Abstention when confidence is low — the feature that brought pilot users back.

94%
citation accuracy
2.1s
p95 response
pilot retention
04
Climate SaaS · ongoing

Java cold starts, quartered

The problem

A Spring API on Lambda where cold starts blew the latency SLO every morning, and the mitigation — provisioned concurrency everywhere — was the largest line on the cloud bill.

What we built

Migrated to GraalVM native-image with reflection configuration generated from the existing test suite, then removed provisioned concurrency. Same code, same framework, different runtime.

3.8s → 0.9s
cold start p95
62%
lambda spend cut
1 sprint
to production
They told us what not to build. That saved us a quarter.
VP Engineering · Series-B logistics platform

Want the longer version?

Ask for the case file — full write-ups with architecture diagrams and the numbers we didn't fit on this page.

Request the case file