AI engineers inside your team until the system is live.
A senior team of forward-deployed AI and ML engineers in London, formerly at Rolls-Royce. We take the work your people do by hand, from calls and photos to documents and systems, and ship it as a working agent, model or pipeline in your environment.
Inbound call
“Water damage in the kitchen, it happened last night.”
2 photos received
Claim form
- Policy number
- 48213
- Excess
- £250
Claim opened
Assessor booked for Thursday. Written to your claims system.
Prototypes are cheap now. Products are not.
A demo takes an afternoon. Making it something your team relies on every day takes real engineers.
The last mile is engineering
A prototype handles the happy path. A product handles bad audio, blurry photos, messy scans, permissions, and the day it breaks at four on a Friday. That is what we do.
Nothing to sell you but the work
No platform, so we are not looking for problems that fit one. We build on whatever models and infrastructure suit you, and you own the code.
Not a vendor's deployment team
Vendors' forward-deployed teams are there to get you onto their product. Our only measure is whether the system works for your people.
What we take on
The manual work your team does every week, and the models and data platforms behind it.
- Voice agents that handle calls
- Intake, bookings, follow-ups and checks, with outcomes written to your systems and a person to hand over to.
- Agents that read documents
- Contracts, statements, scanned forms and reports: extract, classify, check and route, with a person reviewing the exceptions.
- Computer vision for photos and video
- Damage photos, site footage, inspections and product images, described, measured and flagged where your team already works.
- Agents that run a workflow end to end
- Cases, claims and applications across several systems, from intake and evidence to a drafted decision and an audit trail.
- Machine-learning models on your data
- Forecasting, scoring, classification and anomaly detection, trained on your history and monitored so they stay honest.
- Data pipelines and big-data platforms
- Ingest, clean and join your data at scale, on the cloud you already use: the groundwork every model and agent needs.
How it works
From first call to a system your team relies on.
- 1
A 30-minute call
You describe the process. We say honestly whether it is a fit and roughly what it would take.
- 2
A written proposal
Scope, success metrics, timeline and a fixed price. No hourly invoices.
- 3
A two-week trial
Two weeks on your real data before you commit to the full build.
- 4
The engagement
Diagnostic, build and go-live, inside your team. Phase by phase below.
A typical engagement, phase by phase
Usually 12 to 16 weeks
Phase 1
Diagnostic
Map the process with the people who do it, prove the approach on real data, then fix scope and price.
Phase 2
Build
A working system in your environment from week one, reviewed by your team against real cases every week.
Phase 3
Live
Monitoring on, your team trained, exceptions routed to a person. Then a clean handover, or we stay on for support.
A few things we hold to
- We sit with your team, not in a slide deck.
- Working software from the first weeks, not a report at the end.
- Your data stays in your environment.
- Evaluation and monitoring are part of the build.
- No platform. You own the code and can leave at any time.
Is this for you?
We would rather say so on the first call than disappoint you in week eight.
A good fit
- A process your team does by hand every week.
- Finance, legal, healthcare or another data-heavy business, from growing firms to teams inside large organisations.
- You want to see it working in weeks, not read a strategy in months.
What we need from you
- One person on your side who can make decisions without a committee.
- Access to the people who do the work, and to the data, in your environment.
- An hour a week to review real output and tell us what is wrong.
- A view on who runs the system after go-live: your team, or us on support.
Who we are
We met building AI and machine-learning systems at Rolls-Royce, then went on to found products of our own. Full backgrounds on the first call.
- More than eight years each inside large organisations: aerospace, pharmaceuticals, financial services and telecom.
- Founders as well as engineers, with products and open-source tools that teams run in production.
- Led AI architecture for a Google Cloud Premier Partner, and sat on the client side of vendor pitches, so we know what a demo hides.
- Computer vision, NLP, big-data platforms and MLOps by background, long before agents.
The people you meet on the first call are the people who build the system.
Work we have done
Described without naming the client or employer. We walk through each one, screens and code, on the first call.
0%+
accuracy pulling tables and key-value data from scanned documents
0%
cut in prototype-to-production time for an enterprise delivery team
0+
GitHub stars across the voice-agent device and the research assistant we founded
0+ yrs
average experience building ML and AI systems inside large organisations
Inside organisations
- Document AI at a FTSE 100 aerospace group
- Computer vision, OCR and NLP pulling tables and key-value data from scanned documents at over 90% accuracy, for internal teams and pharmaceutical customers. Plus an end-to-end pipeline that analyses annual reports, front end included.
- Clinical-trial data platform for a London biotech
- Built on Google Cloud and published by Google as a customer case study, with drug-discovery pipelines on AlphaFold2 running alongside it.
- Agent platforms for the clients of a Google Cloud Premier Partner
- Multi-agent workflows with persistent memory and vector retrieval for enterprise clients. Reusable evaluation harnesses and deployment scaffolds cut the team's prototype-to-production time by about 30%.
Products we have founded
- Starmoon AI
- A voice agent in hardware: a conversational device small enough to sit inside a toy, speech in and out on a tiny board, plus a companion app. Open source, and used for comfort, such as helping a child through nerves before a medical appointment. View on GitHub
- IncarnaMind
- An open-source research assistant that answers questions across a whole library of PDFs and text files, with the source open beside the chat, saved sessions, and hosted or local models. Teams run it in production. View on GitHub
- Retapp
- Intake and pricing for electronics refurbishers. Devices are catalogued from web and mobile apps, graded, priced against live and sold listings in the UK, France, the US and Germany, then approved for sale or exported in bulk.
Questions we get asked
- Agents that take calls, read photos and video, and work through documents and systems. Underneath, models trained on your data, computer vision and pipelines at scale. All built with evaluation, monitoring, permissions and a clean hand-off to a person.
- Inside your team, with your data and systems, rather than in an office building a demo. Software firms call this forward-deployed engineering. It is our whole business.
- Hosted models, or open-weight models on your own infrastructure, on the cloud you already use. No platform to steer you towards, and you own the code.
- A fixed price for the diagnostic, then a fixed price for the build, agreed once we have seen the real data. No hourly invoices.
- It ships with evaluation and monitoring, and your team is trained to run it. Take it over entirely, or keep us on for support.
- Any size. What matters is a process done by hand, an owner who can decide, and data we can access.
Tell us about the work your team does by hand.
Thirty minutes. We will say honestly whether it is a fit.