Past the demo. We ship retrieval-augmented systems, multi-step agents, fine-tuned models and automation pipelines that survive contact with real users, real data, and real cost-per-token spreadsheets.
Median eval pass rate before launch: 94%. Median cost-per-task drop after our optimisation pass: 6.4×.
Six recurring archetypes cover most of what we get asked to build. Pick the one closest to yours, or describe your own — we've probably done a cousin of it.
Search, chat, summarisation, citation. Built on the boring parts (chunking, hybrid retrieval, reranking, evals) — not just an OpenAI call wrapped in a sycophantic prompt.
Multi-step agents that read, plan, and act on real systems. Tool use, browser automation, human-in-the-loop checkpoints, full audit trail.
When prompting plateaus. SFT, DPO, RLAIF on the open-weight models that fit your task — and the eval harness that proves it works.
Slack-native, IDE-native, dashboard-native. The copilot your team actually uses because it knows your wiki, your issues, your codebase.
Invoices, contracts, claims, KYC packets. We replace the form-shop with structured extraction + verification + human escalation.
Real-time voice agents, vision QA, multimodal search. Built with eval-driven discipline — not vibe-checked.
Every engagement ends the same way: a production system, a team that can run it, and a written record of every decision we argued about.
A deliberately conservative core. We swap in sharp tools when the problem requires them, never because they're trending.
Pick the one that maps to where you are. Or hop between them — most clients start at one and end at another.
Architecture review, prototype of the riskiest interaction, and a costed roadmap. You walk away with a plan whether or not we build it.
Senior pod of 3–5: tech lead, full-stack, design, QA. Two-week cycles, demo every Friday, you own the repo from week one.
Once it's live. A reduced pod stays in your stand-ups, owns the on-call rota, and ships the roadmap quarter-over-quarter.
If yours isn't here, just ask. We respond to most enquiries within a working day.
Yes — most of our work is alongside in-house teams. We default to your repo, your code review, your standup. Where we run our own pods, we still pair-program weekly so the knowledge transfers from day one rather than at handover.
Frequently. We start with a one-week audit: code, infra, team, runway. Output is a frank document — what to keep, what to throw out, what to rewrite — and a costed path to stable. About a third of those projects we then take on. The rest we hand back with a plan they can execute themselves.
You own everything we write — code, designs, infra, docs — assigned on first commit. We keep no kill-switches and no exclusivity clauses. Our standard MSA is six pages and we'll happily redline yours.
For well-shaped, mostly-known scopes — yes, we offer fixed-fee with milestone billing. For genuinely exploratory work we recommend time-and-materials with a hard cap, because anyone who quotes you a fixed price for an unknown is either lying or building you something cheap.
A two-week discovery sprint. Below that we're not adding much value over a good freelancer, and we'd rather refer you to one of the four we trust than do it badly ourselves.