Work

Four projects, described the way we would in a review

Not logos and adjectives. What the business was actually stuck on, the call we made, and the number that moved as a result, including where the answer was to build less.

60+
Products shipped: 94%
Client retention: 11
Time zones: 12 days
Median time to staff

Selected work

Projects, constraints and what actually changed

Client names are held under NDA. The problems, the decisions and the numbers are not, and we will walk through any of these in detail on a call.

HealthcareMulti-site specialty clinic group

One patient record across eleven clinics on four different systems

The problem

Eleven acquired clinics ran four EHRs between them. Front-desk staff were rekeying demographics by hand, and duplicate records were causing scheduling errors serious enough to reach the board.

What we did

We built a FHIR-based integration layer with a deterministic-then-probabilistic patient matcher, migrated two of the four systems onto a hardened OpenEMR deployment, and left the remaining vendors in place behind the same interface. Every match above the review threshold is logged and reversible.

FHIR R4
HL7 v2
OpenEMR
Python
PostgreSQL
AWS

11 to 1
Patient record systems of truth: 97.4%
Auto-match rate on migration: 6.5 hrs
Front-desk hours saved per site, weekly

Duration: 7 months
Team: 5 engineers, 1 clinical analyst

FintechB2B payments platform, New York

Rebuilding a ledger that had drifted $180k out of balance

The problem

Balances were computed from a mutable transactions table with no audit history. Month-end close took nine days and finance had lost confidence in the numbers, which was becoming a problem in diligence.

What we did

We introduced an append-only double-entry ledger alongside the existing system, dual-wrote for six weeks while reconciling continuously, then cut over once the two agreed for twenty-one consecutive days. Historical drift was traced to three specific refund paths and corrected with documented adjusting entries.

Stripe
Node.js
TypeScript
PostgreSQL
Terraform

9 days to 4 hrs
Month-end close: $0
Unexplained variance since cutover: 21 days
Parallel-run agreement before cutover

Duration: 5 months
Team: 4 engineers

AIHealthcare data company

A retrieval agent clinicians will actually put their name to

The problem

An internal LLM prototype answered clinical questions fluently and cited nothing. Clinical leadership would not approve it, and per-query cost was rising faster than usage.

What we did

We rebuilt it on LangGraph with mandatory source attribution. The agent cannot return an answer without passing citations, and refuses rather than guessing. We wrote an evaluation set of 640 clinician-labeled questions, then used it to justify routing two thirds of traffic to a smaller model.

LangGraph
LangChain
pgvector
Python
Anthropic
OpenAI

94%
Answers passing clinician review: 58%
Reduction in cost per query: 640
Labeled questions in the eval set

Duration: 4 months
Team: 3 engineers, 1 ML engineer

AgricultureRow-crop agronomy cooperative

Field software that works where there is no signal

The problem

Agronomists were recording scouting observations on paper because the existing web app was unusable beyond the edge of coverage. Data reached the office days late, often with gaps.

What we did

An offline-first mobile application with conflict-aware sync, backed by a pipeline that joins the field observations to satellite NDVI and local weather. Sync is designed around the reality that a device may be offline for a full working day and then reconnect all at once.

React Native
SQLite
Python
PostGIS
GCP

340k
Acres under active management: 3 days to same day
Observation to office: 0.4%
Sync conflicts requiring review

Duration: 6 months
Team: 4 engineers

Companies we have worked with

Sully AI

Pony.ai

Meridian Health

Northlake Capital

Vantage Pay

Alder Diagnostics

Brightpath Learning

Cardinal Bancorp

Tillage Systems

Axle Motion

Halcyon Care

Ledgerline

Summit Scholars

Rowan Robotics

Sully AI

Pony.ai

Meridian Health

Northlake Capital

Vantage Pay

Alder Diagnostics

Brightpath Learning

Cardinal Bancorp

Tillage Systems

Axle Motion

Halcyon Care

Ledgerline

Summit Scholars

Rowan Robotics

In their words

What clients say when the project is over

They spent the first two weeks telling us the project we asked for was the wrong one, and were right. The version we actually built was smaller, cheaper and is still running.

VP EngineeringMulti-site specialty clinic groupName withheld under NDA

The ledger cutover was the least dramatic infrastructure change we have ever done. Twenty-one days of parallel running and then it was just done.

Chief Technology OfficerB2B payments platform, New YorkName withheld under NDA

We had been quoted a rewrite by two other firms. ML Stack Labs carved it up and shipped the first piece in five weeks without taking the system down.

Director of TechnologyK-12 learning platformName withheld under NDA

Tell us what is on fire.

Bring us a stalled project, a system nobody wants to touch, or a product you need built properly the first time. We will tell you honestly whether we are the right team for it.

Book a scoping call Contact us

45 minutes, no cost, no deck
You speak to an engineer, not a salesperson
We say no if it is not a fit