Applied AI Research & Development
Applied AI research for problems without a known solution. We co-author with UC Berkeley, Harvard and UCL, then deploy what works into your stack.
Some problems do not have a known solution yet. The data is unusual, the constraints are tight, or the published methods do not survive contact with production. That is where our Research & Development practice works: we take open problems, run the research needed to solve them, and turn the result into a system your team can operate.
What we take on
- Novel architectures for problems where standard models underperform, such as diffusion-based reconstruction or multiscale spatial-spectral networks.
- Performance research at the systems level: kernels, quantization and parallelism strategies that change what is feasible on given hardware.
- Feasibility studies that answer, with evidence, whether a target accuracy, latency or cost is achievable before you commit to a full build.
- Research-to-production transfer: taking a promising paper or internal prototype and making it reliable, measurable and deployable.
How we run research so it ships
Start with the decision, not the paper
Every project begins with the operational question the research must answer and the metric that would change your plans. That keeps the work pointed at deployment from day one.
Structure uncertainty
We break the work into experiments with explicit success criteria and checkpoints. If a direction is not working, you hear about it at the checkpoint, along with the evidence and the alternatives.
Build production habits into research
Reproducible data pipelines, versioned experiments and evaluation harnesses are set up at the start, not retrofitted. When a result is promising, the path to production is short. The same harness that scored an experiment becomes the regression test for the deployed system, so nothing about the result has to be rebuilt to ship it.
Leave capability behind
Your engineers work alongside ours for the whole project, so the knowledge stays in-house when we step back.
Research that reached production
- DeepSeek R1 FP4 Inference World Record: a world-record 303 tokens per second on DeepSeek R1 in FP4, set with Avian.io and NVIDIA on a single DGX Blackwell node and verified by Artificial Analysis.
- CT Scan Super-Resolution & Denoising: a diffusion-based architecture for simultaneous 4× super-resolution and denoising, enabling diagnostic-quality CT from 75% reduced radiation exposure.
- Virtual Try-On Diffusion Model: a virtual try-on diffusion model built from scratch for Wearit, trained in distributed A100/H100 runs on the Jean Zay supercomputer under a GENCI compute grant, on 3M+ custom-labeled fashion images, with production inference under 2 seconds.
- Identity Verification Engine: print-technique forensics and multi-spectral authenticity analysis for Oneex, developed with a former IRCGN document-fraud expert and running on CPU-only edge hardware.
- Hyperspectral Heavy Metal Detection: custom spectral unmixing and a satellite imagery pipeline reaching 94.2% detection accuracy. The method is described in our technical article on multiscale spatial deep learning.
Academic and industry grounding
We co-author with UC Berkeley, Harvard and UCL, and our team has operator experience at NVIDIA, Siemens Healthineers and JP Morgan. That combination is deliberate: academic collaborators keep us close to new methods, and operator experience keeps us honest about what production demands.
If your bottleneck is compute, see R&D & Frontier Compute. If you have an open problem and want a candid view on whether it is solvable, start a conversation.
Frequently asked questions
How is an R&D engagement different from a standard build?
In a standard build the approach is known and the work is execution. In an R&D engagement the approach itself is uncertain, so we structure the work around experiments and decision points, and we tell you early when a direction is not paying off.
Who owns the results and intellectual property?
Ownership terms are agreed per engagement before work begins. Our default is that you leave with the code, models and documentation your team needs to run and extend the system.
Do you publish the work?
Only where it serves the work and you agree. Our team has published peer-reviewed medical imaging research with Siemens Healthineers and UCL, and built the inference stack behind a DeepSeek R1 FP4 world record on NVIDIA Blackwell, set with Avian.io and NVIDIA.
Who is on the team?
A small, senior team of AI researchers and engineers, including PhDs and research alumni, based in San Francisco and Dubai. We co-author with UC Berkeley, Harvard and UCL.
Selected work
- Computer VisionAI Identity Document Verification Engine for OneexDesigned and built the AI engine of Oneex's identity-verification platform, which processes over 10 million identity documents a year in airports, banking, defence and sovereign infrastructure. OCR, multi-spectral authenticity analysis, print-technique forensics and biometrics, all running on CPU-only edge hardware.
- Remote SensingHyperspectral Heavy Metal DetectionBuilt a complete pipeline for processing hyperspectral satellite imagery to detect and quantify heavy metal contamination in soil across mining regions. Achieved 94.2% detection accuracy using custom spectral unmixing algorithms.
- Medical ImagingCT Scan Super-Resolution & DenoisingDeveloped a novel diffusion-based architecture for simultaneous super-resolution (4x) and denoising of low-dose CT scans, enabling diagnostic-quality images from 75% reduced radiation exposure.
- LLM OptimizationDeepSeek R1 Inference World Record on NVIDIA BlackwellWith Avian.io and NVIDIA, our team built the inference stack behind a DeepSeek R1 world record: 303 output tokens per second in FP4 on a single NVIDIA DGX Blackwell node, independently benchmarked by Artificial Analysis.
- Generative AIVirtual Try-On Diffusion Model for WearitArchitected Wearit's core generative AI pipeline for virtual clothing try-on, from R&D to real-time production: a diffusion model built from scratch, trained on the Jean Zay supercomputer on 3M+ labeled fashion images, serving results in under 2 seconds.
- Document AIIndustrial OCR Engine Audit & Roadmap for TessiAdvised the Technical Direction of Tessi, a European leader in document processing, on the evolution of their industrial OCR engine: a benchmark against state-of-the-art models, an end-to-end architecture audit and a technical roadmap for the next-generation platform.
Industries we apply it in
- IndustryR&D & Frontier ComputeGPU performance engineering for teams limited by throughput or cost: custom CUDA kernels, quantization, distributed inference and training infrastructure. We contribute to open source where it helps the work.
- IndustryHealthcare & Life SciencesMedical imaging, real-time ultrasound AI and diagnostic systems built to hospital reliability standards. We have built real-time endometriosis detection on ultrasound video, CT super-resolution and registration pipelines with Siemens Healthineers and academic partners.
- IndustryIndustrial & Public SectorHyperspectral remote sensing, geospatial intelligence and edge vision for industrial operators and the public sector. Our models run on satellite data and on edge hardware at remote sites with little bandwidth.
Have a research & development problem worth solving properly?
Book an intro call