What we build

AI agents that do real work

Agents that plan, act and report back — not chatbots. Research agents, internal assistants, and workflow agents that touch your real systems.

What you get

• An agent that plans its own steps and runs them against your systems
• A live view of what it is doing, so a run is never a black box
• Failure handling that degrades instead of stopping: a partial result, flagged
• Source code, deployment and a walkthrough for your team

Who it’s for. Teams drowning in manual research, triage or data gathering.

Evidence. SourceTrail plans 10+ questions, searches, reads and verifies every claim before it writes.

Read how it works: SourceTrail

Browser and workflow automation at scale

Automating what only exists behind a login or a UI — portals with no API, form flows, data extraction — running reliably on real infrastructure, with a human able to step in when a page breaks.

What you get

• An HTTP API that starts a task and returns a live URL to watch it
• Human handoff: take control mid-run, then hand it back
• Credential handling where secrets never reach the model or the logs
• Containerised deployment that scales with the queue

Who it’s for. Ops teams doing repetitive portal work, and companies whose vendors have no API.

Evidence. Francurial returns a task ID and a signed live session URL in under 100ms, and pauses for a human on a CAPTCHA.

Read how it works: Francurial

Real-time AI applications

Desktop and web apps that process speech, video and screen content live — transcription, vision, monitoring, analysis — and produce something useful the moment it happens.

What you get

• A capture and inference pipeline built for latency, not batch throughput
• Models loaded once and kept warm, with a measured latency budget
• Clean degradation: one model failing switches off one feature, not the app
• A reviewable record of what the system saw and when

Who it’s for. Products where the AI has to keep up with a conversation or a stream.

Evidence. Sherlock AI tracks participants across frames live; Gandiva AI transcribes and reads the screen as you talk.

Read how it works: Sherlock AI, Gandiva AI

LLM integration and retrieval systems

Putting your own documents, data and processes behind a model so answers are grounded in your business rather than the open internet.

What you get

• A retrieval pipeline over your content, with the chunking and ranking tuned to it
• Citations that resolve to a real stored source, checked before the answer ships
• A prompt and context budget that keeps cost flat as usage grows
• An evaluation set, so a change can be shown to be an improvement

Who it’s for. Companies with a knowledge base, a support load, or a document problem.

Evidence. SourceTrail’s retrieval and memory layer makes a fabricated citation structurally impossible; Gandiva AI answers from a local vector store.

Read how it works: SourceTrail, Gandiva AI

AI/ML engineering

Training, fine-tuning and serving models on your own data, for when a prompt is not the answer: classification, extraction, vision, speech, forecasting and ranking.

What you get

• A labelled dataset and a held-out evaluation set you keep
• A measured baseline before anything is trained, so improvement is a number
• A trained or fine-tuned model with accuracy, latency and cost recorded
• An inference service, monitoring, and a documented retraining path

Who it’s for. Teams with proprietary data where an off-the-shelf model does not do the job well enough.

We check first whether an off-the-shelf model already solves it, and tell you if it does. The cheapest answer that works wins.

Evidence. The model pipelines inside Sherlock AI (face embeddings, OCR, speech) and Gandiva AI (local speech recognition with a domain-vocabulary correction pass) are on the critical path, not behind an API wrapper.

Read how it works: Sherlock AI, Gandiva AI

Full-stack product engineering

The product around the AI, or the product on its own: API, database, auth, dashboard, background jobs, deployment and the CI that keeps it honest.

What you get

• A typed API and a schema you can read
• The web app, admin, authentication, roles and permissions
• Background jobs and queues where the work is slow
• Containerised deploy, CI/CD, tests on every push, and a handover walkthrough

Who it’s for. Founders taking a prototype to production, and teams whose AI feature has no product around it yet.

Evidence. SourceTrail (FastAPI, PostgreSQL, Qdrant, Electron, 164 tests) and Francurial (Redis Streams, WebSockets, Kubernetes autoscaling) are complete systems with queues, storage and deployment.

Read how it works: SourceTrail, Francurial

Blockchain and smart contracts

On-chain systems where the ledger is the point: tokens, escrow and programmable payments, provenance, wallet authentication, and the indexing and app layer that makes any of it usable.

What you get

• Contracts with a full test suite, including the adversarial cases
• Deployment scripts, a testnet deployment before mainnet, and a gas report
• An indexer or subgraph, so the frontend is not reading chain state directly
• Wallet integration, and ownership and upgrade rights transferred to you

Who it’s for. Teams that need settlement, provenance or programmable payments, and teams already on-chain who need the application layer.

For anything holding real value we arrange an independent third-party audit before mainnet, as a line in the proposal. We do not self-certify our own contracts as secure.

Evidence. Nothing public yet on this line — our open-source work is AI, backend and desktop. We would rather say that than point you at something adjacent. The scoping sprint is the low-risk way to test the fit.

Mobile app development

iOS and Android apps, including the ones where the hard part is on-device: camera, microphone, offline-first and background sync.

What you get

• The app in React Native, or native Swift and Kotlin where the platform demands it
• API integration, local storage, and sync that survives a bad connection
• Push notifications, and submission to both stores including the review paperwork
• The release and signing process handed to your team, so you are not locked to us

Who it’s for. Teams that need their product in users' hands, and anyone whose web or desktop AI feature needs a mobile client.

Evidence. Nothing public yet on this line — our open-source work is AI, backend and desktop. We would rather say that than point you at something adjacent. The scoping sprint is the low-risk way to test the fit.