Wanx AI is a research lab that ships. We train and fine-tune ML models, build the data pipelines and agentic systems that run them at real scale, and turn the resulting research into production software — for a fraction of what naive AI infrastructure costs. Remote-first, working with clients globally.
Same task, two architectures — cost per 1,000,000 files processed
From AIDAW, our agentic media-processing engine — details below.
Three capabilities, one team — software, models, and the data layer underneath them.
Production systems, not prototypes — backend architecture, databases, integrations, and the operational tooling that makes software usable by a real team on day one.
Proof: GarageOS — multi-branch operations platform. FastAPI, PostgreSQL, Next.js, M-Pesa payments built in.
When a general-purpose model isn't accurate enough on your data, we adapt one that is — LoRA fine-tuning, evaluation harnesses, and the failure analysis that tells you exactly where a model still breaks.
Proof: R.O.A.D. — Qwen3-VL fine-tuned for historical handwriting recognition on degraded archival records. Hydra II — a production text-to-music model.
Multi-agent systems that process real volume without the context bill exploding. We design the data layer first, then build agents that only look at what they actually need to.
Proof: AIDAW — 2.5M files indexed in ~3 minutes, 70+ tools across 5 agents, 4 LLM providers.
Systems we designed, built, and shipped — with the numbers to back them up.
A 5-agent system (Supervisor, Reaper, FFmpeg, Search, Data — 70+ tools) that lets studios process millions of media files through natural language instead of manual pipelines.
A trust-infrastructure platform for multi-branch auto-repair chains — real-time customer transparency, M-Pesa payments, and a weighted Trust Score computed per job.
A fine-tuning and evaluation pipeline for historical document handwriting recognition, built on Qwen3-VL and tuned for degraded archival records.
Two engagements, two very different problems — the pattern underneath is the same research-to-production pipeline.
A Nairobi-based NGO working toward an Africa where children and youth with disabilities can realise their full potential. We built an early-childhood-development content generation app for them — combining generated background music with unique synthesized narrator voices to produce accessible storytelling content.
An AI music company building synthetic training datasets and licensing infrastructure for the music industry. We build and maintain Hydra II, their production text-to-music generation model — from research and training through deployment and ongoing evaluation.
The same four steps on every engagement, in order.
We turn your requirements into a precise technical specification before any code is written.
Engineers — not just prompts — implement the system, with evaluation criteria defined upfront.
Every model and pipeline ships with a measurable evaluation harness, not a demo that happened to work once.
We hand off with monitoring and documentation in place, or stay on to run it.
Wanx AI is a remote-first research lab founded in Nairobi, working with clients globally. The same rigor we bring to novel research — physics-informed neural networks, neural operators, custom multi-agent RL environments — carries into every client engagement. We're a small team that ships production systems ourselves — no account managers standing between you and the engineers building your system.
Send a note about the problem you're solving — software, a model, or a pipeline that needs to run at real scale — and we'll get back to you.