Custom agents
An agent made for one job.
Generic agents run on trillion-parameter models built to do everything. Most business work is one job done many times. We fine-tune or distill a small model for that job, connect it to your tools, measure it against the generic option on your own data, and hand it over to run on your infrastructure.
Built on one 8 GB laptop GPU
- Razin OCROur own text detector and line reader for Arabic and English pages, trained on licensed and synthetic pages.
- globalmmA 16.8M-parameter connector gives a text-only Llama-3.2-1B grounded vision: 86% on held-out "where is the X?" boxes, against 91% for Qwen3-VL-2B.
- argusSixteen live YOLO video streams on one GPU: a 300-frame run went from 27.0 s to 7.0 s, about 770 frames a second.
The write-ups carry the method and every number.
Why a smaller model.
Cheaper per task
A model with a few billion parameters costs far less to run per call, and on your own hardware there's no per-token bill.
Private
It runs on your servers or your devices, so documents and customer data don't leave them.
Predictable
It's trained and tested on the one job it does, so how well it does that job is measured, not hoped for.
Yours
You keep the weights, the code and the evaluation set. No lock-in to us or to a model vendor.
What we build.
Fine-tuned and distilled agents
We start from an open model and fine-tune it on your examples, or distill a larger model's behaviour on your task into it, then put it inside an agent that calls your tools.
Built on how we train Razin OCR and globalmm's connectorsDocument AI, Arabic included
OCR and extraction for documents in Arabic, English and mixed scripts, evaluated on your own pages before anything goes to production.
Built on our work on Razin OCRAgents on the tools you already use
MCP servers, and the web apps you already run turned into tools agents can call safely. Coding agents on your own computers, directed from a phone.
Built on what we learned building Deputy and LinkOn-device and private AI
Choose, quantize and run models on phones, laptops and your own servers, with GPU and NPU acceleration and offline-first apps.
Built on what we learned shipping OrbHow an engagement runs.
- 01
Scope
One call and a written statement of the job, with examples of it done well.
- 02
Baseline
Measure the generic option, such as a frontier model behind an API, on your own data. That's the number to beat.
- 03
Train
Fine-tune or distill a small model for the job, build the agent around it, and measure it the same way.
- 04
Hand over
Weights, code and evaluation reports, running on your infrastructure.
What you keep
- Model weights and code, in your repository.
- Evaluation reports you can rerun yourself.
- Documentation written for the people who'll maintain it.
- No lock-in to us or to a single model vendor.
Tell us about the job.
A short description is enough to start. We'll reply with whether a smaller model can do it, and how we'd measure that.