Raw inference.
Real engineering.
No wrappers. No compromises. No cloud dependencies.
10× Faster
Sub-millisecond inference. No API roundtrips. Your model runs on your hardware, on your terms.
100% Private
Your data never leaves your infrastructure. AES-256. Complete control over every byte.
You Own It
No black boxes. You see every decision. You get the weights, the code, everything. No licensing fees.
Autonomase.
A reinforcement learning agent for autonomous web navigation. Gets smarter with every site it visits — and carries that knowledge forward.
Most scrapers start from zero every time. Autonomase doesn't. It navigates websites autonomously using a neurosymbolic RL policy trained on symbolic DOM structure — not raw HTML noise. Every successful navigation updates its beliefs. Those beliefs transfer to structurally similar sites.
First visit: broad exploration. Tenth visit: near-deterministic. Learned from AutoTrader? Motors.co.uk bootstraps in a fraction of the time. The more it runs, the better it gets — and competitors can't replicate that without the same interaction history.
Agent loop
Observe
DOM snapshot → symbolic parser extracts buttons, links, fields, ARIA labels, URL patterns
Decide
RL policy acts on clean symbolic state — not raw HTML — massively reducing state space
Act
Click / type / scroll via Playwright. Symbolic check gates every action before execution
Update
Successful trajectory stored to episodic memory. Posterior updated: 'links with this class → listing data'
Transfer
Prior carries to structurally similar sites. Car sites share DOM patterns. Prior bootstraps faster each time
Reward signal
make, model, year, mileage, price
clicked into listing from search
episode terminated
hash match → loop prevention
Why PPO
On-policy. Handles diverse and changing action spaces across different sites naturally — which off-policy methods struggle with. Episodic memory adds sample efficiency without needing a world model.
Model-based (e.g. DreamerV3) would need retraining per site — different transition dynamics every time. Model-free generalises across diverse environments by design.
The moat
Generic scrapers start from zero on every site. Autonomase accumulates cross-site experience. The more sites it navigates, the better its priors, the faster it converges on new ones.
Competitors cannot replicate this without the same interaction history. The data is the moat. The agent gets it.
First target
UK used car listings — concrete, evaluable: did the agent extract make, model, year, mileage, price, location?
Things we're building.
Powered by Autonomase. Built for people who want better.
Trailset.
Job market intelligence built on real data — not aggregated job boards. Know which companies actually hire, how they move, and whether you're a genuine fit before you waste your time applying.
Powered by Autonomase crawling hundreds of UK career sources. No direct competitor. Biodiversity of data is the moat.
Gym × RL.
A reinforcement learning agent that learns your body — your structure, your goals, your recovery. Built for every body type, not the average.
The kind of thing that comes from actually being in the gym and knowing RL from first principles. Social accountability layer built in — progress you can share, not just stats you ignore.
RLFoundry.
A clean-room RL library built from first principles. PPO, safe RL, constraint satisfaction — implemented properly.
Built as part of ongoing research into safe reinforcement learning and cross-domain constraint transfer. Currently implementing PPO → PPO-Lagrangian → CPO across structurally distinct applied environments — healthcare, energy, smart grids, robotics.
No wrappers around existing libraries. Everything from scratch, against CleanRL as specification. Infrastructure stack: Hydra, Weights & Biases, Slurm/Apptainer on HPC.
Research targets
Select applied ML work.
We take on a small number of projects. Edge inference, safe RL, custom training pipelines.
Computer vision, NLP, forecasting — built and deployed on your infrastructure. We don't wrap APIs. We train real models. We optimise for your hardware.
Philosophy
We don't believe in wrappers. We believe in real engineering. If you need someone to bolt GPT-4 onto a chatbot, we're not the right fit. If you're solving something genuinely hard on real hardware — let's talk.
Founded on the principle that AI shouldn't be locked behind expensive APIs. Open tools. Open access. Real independence.
0%
Local inference
0+
Research domains
0 cloud deps
Hard dependencies
0+
Trailset data sources
Built with conviction.
No gatekeeping
AI shouldn't be locked behind expensive APIs. Open tools. Open access. Real independence. If you can run it locally, you should.
Real engineering
We solve hard problems on real hardware. No handholding, no shortcuts. The kind of work that holds up under scrutiny.
Your control
You own the models. You own the data. No lock-in, no vendor dependency. Complete freedom — by design, not as an afterthought.
Safety and ethics
by default.
Responsible AI
Bias detection, fairness monitoring, and explainability from day one. No black boxes. No hidden assumptions.
Data governance
Your data stays yours. No retraining on your data without explicit consent. Full transparency on what's stored and how long we keep it.
Human oversight
AI that supports decision-making, not replaces it. Built with override capabilities. Humans stay in control. Always.