What we build
From the model to the product it ships inside
Nine services, three groups. Most projects touch more than one.
Intelligence
Getting a model to do something useful with your data.
- AI Automation
- AI Modeling
- AI Chatbots & Systems
Systems
The engineering that makes it run the same way every time.
- Pipelines
- Systems & Engines
- Software Process
Product
What the people you are building for actually touch.
- Websites
- Arts
- Consulting
Selected work
Things we have actually shipped
Engines, models and platforms — shipped for competitions, clients and ourselves.
ChaosX Engine
In development
A C++ game engine built from the ground up — custom application and entry-point architecture, event system, and logging, targeting Windows and Linux.
- C++
- premake5
- Cross-platform
GravitySort
Private4.6x cub::DeviceRadixSort on low-cardinality data, parity on uniform · N=10⁹, A100-80GB
A distribution sort for int32 keys, CPU and GPU. A gateway samples the input and routes it to a pipeline chosen for its distribution — counting and filling when values repeat, exiting after one pass when data already arrives ordered — so each route skips the work it does not need.
- CUDA
- A100
- Research
Energy Degradation Prediction
1st place — IEEE Jordan AI Modeling Hackathon 2.0 · RMSLE 0.19877
Predicting an energy degradation index across 42 channels under train/test distributional shift, blending four tabular foundation models.
- TabPFN
- Nori-30M
- TabFM
Chest X-Ray Bench
Best ensemble 0.9174 AUROC on test500
A controlled comparison of 43 chest X-ray models on CheXpert, separating the design choices that actually move scores from the ones that are just noise.
- CheXpert
- ConvNeXt
- Medical imaging
Sign Language Bridge
Fine-tuning Qwen3-VL-2B-Instruct for continuous ASL to English translation, with multi-tier LoRA, staged OpenASL and How2Sign training, and a MediaPipe preprocessing pipeline.
- Qwen3-VL
- LoRA
- MediaPipe
Health-Navigator
An agentic medical assistant that reasons over patient data through multi-agent orchestration, combining pre-trained ML models with relational and vector databases.
- Multi-agent
- Vector DB
- Jupyter
Team-Finder
Security-hardened team-matching platform that pairs students by skill similarity and verifies members against university email domains.
- Next.js
- FastAPI
- Supabase
Video Captioning
Samples frames with ffmpeg and sends them to vision models for captions in four voices, then runs an ensemble with a judge model picking the best result.
- VLM ensemble
- ffmpeg
- Python
RSNA Knee Abnormality Detection
Detecting knee abnormalities from RSNA medical imaging data.
- Medical imaging
- Deep learning
How we work
What happens after you email us
Four steps, in order. You see something running before the end.
- 01
Scope
We tell you what it takes, before any money moves.
- 02
Prototype
The riskiest part first, so it fails cheap if it is going to.
- 03
Build
Shipped in slices you can see running, not one delivery at the end.
- 04
Hand over
Documented and readable, so it is yours to maintain without us.
Tell Us What You Are Building
Describe the idea in your own words and we will work out which of the nine it needs.