Our work

Six projects, and what they actually changed.

Every number below was measured after launch, by the client, in production. Where a project was hard, the hard part is written down.

AIHelios Bank · 2025

Helios Orbit

An AI analyst that reads every filing before the market opens.

The problem

Helios analysts lost the first three hours of every day to document triage. Off-the-shelf assistants hallucinated numbers — unusable in a regulated desk where a wrong basis point is a compliance event.

What we did

We built a retrieval layer over 40 years of filings with span-level citations, then an agent graph that plans, retrieves, cross-checks and refuses. Every answer carries provenance; anything under the confidence floor escalates to a human instead of guessing.

The outcome

Morning briefs land at 06:00 with zero unsourced claims across 18 months of production. The analyst desk redeployed 3 hours a day from triage to judgment.

  • 3hSaved per analyst daily
  • 0Unsourced claims in prod
  • 40yrCorpus indexed
  • Next.js
  • Python
  • OpenAI
  • PostgreSQL
  • AWS
GamesKite Studios · 2025

Ashfall

A 64-player extraction shooter with rollback netcode.

The problem

Kite had a beloved prototype and eleven months to a locked launch window, across three platforms with three certification regimes and no netcode.

What we did

We rebuilt the simulation deterministically for rollback, cut a 40ms frame budget into 16ms with GPU instancing and Nanite discipline, and ran certification as a parallel track from month one rather than a wall at the end.

The outcome

Shipped on time on all three platforms. 1.2 million players in the first fortnight and a 91% positive rating that held through the first content season.

  • 1.2MPlayers in 14 days
  • 91%Positive reviews
  • 16msLocked frame budget
  • Unreal
  • C++
  • AWS
  • Kubernetes
  • Rust
EnterpriseVantatek · 2024

Vantatek Atlas

A 22-year-old ERP replaced without a single hour of downtime.

The problem

The system of record was COBOL, undocumented, and load-bearing for every warehouse Vantatek operated. Two previous rewrite attempts had been abandoned mid-flight.

What we did

No rewrite. We put an event gateway in front of the mainframe, moved one bounded context at a time behind it, and dual-wrote until the new path proved itself under real production load. Rollback stayed one flag away for the entire programme.

The outcome

Fourteen months, zero downtime, zero lost orders. Order throughput up 3.4× and a codebase the client's own team now ships to weekly.

  • 0Minutes of downtime
  • 3.4×Throughput increase
  • 400kOrders per day
  • TypeScript
  • Go
  • PostgreSQL
  • Kubernetes
  • Azure
MobileOrbital Health · 2025

Orbital

A clinical companion that works where the signal doesn't.

The problem

Clinicians worked hours from coverage. Existing tooling assumed connectivity and silently lost records — the worst possible failure mode in a patient chart.

What we did

We made offline the default rather than the fallback: a CRDT-backed local store, deterministic merge on reconnect, and an audit trail that satisfies HIPAA without asking clinicians to think about sync at all.

The outcome

Deployed to 6,000 clinicians. Not one record lost in 11 months, and charting time per patient dropped by 38%.

  • 6,000Clinicians onboard
  • 38%Less charting time
  • 100%Record integrity
  • Flutter
  • Swift
  • Node.js
  • PostgreSQL
  • Google Cloud
XRNorthwind Robotics · 2024

Northwind Forge

Factory training in VR that cut incidents by half.

The problem

Training on live cells meant halting a production line worth $12,000 an hour — so it happened rarely, and incidents kept happening.

What we did

We rebuilt the cell in Unity against the real controller firmware, so the simulated robot fails exactly the way the physical one does. Held at 90fps on-device, because comfort is the difference between a tool used daily and one left in a drawer.

The outcome

Rolled out to 11 plants. Safety incidents fell 52% year over year and time-to-certify a new technician dropped from six weeks to nine days.

  • 52%Fewer incidents
  • 9 daysTo certify, from 6 weeks
  • 90fpsLocked on-device
  • Unity
  • C#
  • Azure
  • Python
CloudMeridian Logistics · 2025

Meridian Signal

Two million events a second, priced to survive a recession.

The problem

The incumbent pipeline buckled at 200k events/sec and the cloud bill grew faster than the fleet did.

What we did

We modelled cost per event before writing infrastructure, then built to that number: Rust ingest at the edge, tiered storage with automatic demotion, and Kafka partitioned by the access pattern rather than by convenience.

The outcome

2.1M events/sec sustained at p99 under 40ms, on 61% less infrastructure spend than the system it replaced.

  • 2.1MEvents per second
  • 61%Lower cloud spend
  • 40msp99 latency
  • Rust
  • Go
  • Kubernetes
  • AWS
  • PostgreSQL