The real kinds of system and architecture we work in — anonymised, with illustrative figures rather than a verified scorecard. Open any one for the problem, the full architecture, and the systems involved.
A health-tech startup needed to pull patient context out of incoming emails the moment they arrived. The hard constraint: patient data could never touch a public AI model.
A law firm runs cases with dozens of documents each — many of them scanned PDFs. Lawyers had no way to ask a question and get an answer grounded in the actual filings.
A client was inspecting structures by eye from photos — slow, inconsistent, and easy to miss a defect.
A client chased payments by hand from a Google Sheet — re-typing the same reminder emails over and over.
A US distributor ran orders in one system and inventory in another, checking them side-by-side by hand, then manually approving, adjusting, and emailing clients and the warehouse.
Connecting to many different EMR systems is a fragmented mess — every vendor speaks its own dialect.
A linear A→B→C is a demo. Real systems look like this: an autoscaling edge, a durable queue, an orchestrator that routes every step to the right kind of worker — and a retrieval layer that decides, per step, whether the answer needs an agent, a tool, or a plain function. Every case is a specialisation of this.
A sub-agent — when a step is multi-step and needs judgment (research a lead, reconcile a claim, draft + revise).
A single external action or lookup — call the EMR, hit a CRM, send an email, run a search.
Plain deterministic code — no LLM. Math, parsing, formatting, a rules check. Cheaper and never wrong.
decision: agent if it reasons · tool if it acts on a system · function if the rules are fixed
The decisions that separate a system that survives production from a demo that doesn't.
We don't force everything through one model. Each step gets the right tool — a retriever, a classifier, an OCR pass, a deterministic script — chosen for the job, not for the hype.
Steps hand off through a durable workflow that checkpoints state, retries on failure, and resumes where it left off instead of restarting from zero.
Near-zero for extraction, matching, and compliance where there's exactly one right answer; higher only where drafting genuinely needs range. Never a single global setting.
Retrieval (RAG / Context7-style) grounds every answer in the client's actual documents and systems — never the model's guess from training data.
Agents pass work to one another and escalate the genuinely ambiguous cases to a person. Nobody — and no model — is left guessing who owns the next step.
Deployed in the client's own Azure or AWS — using their credits where it helps — so sensitive data and the bill both stay with them.
If it isn't, it's close. Tell me the slow, manual, expensive thing — I'll show you the architecture to kill it.
Start a build →