The engineering behind
an AI system that holds up.

Getting something impressive working takes a week. Getting it accurate on your own material, safe with your own data, fast enough at your busiest hour, and provable to an auditor takes the rest. This is the part we do, and these are the six pieces of it.

This is for you if

The six

What we build, and which problem each one solves

Most projects use two or three of these together. A knowledge assistant for a bank, for example, is retrieval plus a private model plus the testing to prove it. They are separate pages because they are separate decisions with separate costs.

Describe what you are trying to build and we will tell you which two or three of these it needs.

Where projects stall

The five reasons a working demonstration never ships

We see the same five, in roughly this order. None of them is about the model, which is why picking a better one rarely helps.

Nobody agreed what correct means

Without a set of real cases and agreed answers, "is it good enough" becomes a matter of whose opinion is loudest, and the project cannot be signed off by anyone.

Where this is handledTesting and evaluation, and it is the first thing we build on every project.

The documents were harder than they looked

Scans, tables and drawings behave nothing like clean text. A specification table read as one line of numbers produces confident answers that are wrong.

Where this is handledRAG and knowledge assistants, in the reading and splitting work.

Legal said no halfway through

The demonstration used a public service and nobody asked where the data went until the security review. The project then waits for months.

Where this is handledPrivate LLM, and settling the data question before the build rather than after.

It could not be connected to anything

Reading is easy. Writing a result back into an ERP, with the right permissions and an audit trail, is where the actual weeks go.

Where this is handledThe integration work, which is what we were doing for eleven years before this.

Nobody owned it after launch

Accuracy slid, the exception queue grew, and the team went back to the old way without telling anyone it had stopped working.

Where this is handledManaged operations and team adoption, on the workflow automation side.

Our stack

The tools, standards and methods we use

A summary across all six. Each page carries the full list for its own area, with a plain explanation of what everything is for.

ModelsHosted where that is allowed, open and self-hosted where it is not. Chosen by testing on your tasks, not by published rankings
Anthropic ClaudeOpenAIAzure OpenAILlamaMistralQwenGemma
Serving and running themYour cloud or ours, with one internal endpoint in front so nothing is tied to a single provider
vLLMModel gatewayDockerKubernetesAzure, AWS and Google CloudAir-gapped installs
RetrievalFinding the right passage before answering, which is where accuracy is usually won or lost
pgvectorQdrantElasticsearchHybrid searchRerankersPermission-aware retrieval
Adapting modelsLight methods that train a small add-on rather than the whole model, so it is affordable and reversible
LoRAQLoRADistillationHugging FacePreference training
Agents and toolsNarrow tools with their own permissions, and the emerging standard for exposing them once to any AI system
Model Context ProtocolTool callingLangGraphTemporalApproval workflows
Testing and observingScored on every change, traced in production, and sampled continuously after release
RAGASDeepEvalpromptfooLangfuseOpenTelemetryPlaywright
SecurityIncluding the attacks specific to AI, which ordinary application testing does not look for
OWASP Top 10 for LLMsPrompt injection testingSingle sign-onPer-tool permissionsRed teaming
Rules and governanceWhich apply depends on your industry and your buyers. We tell you which, before the build
EU AI ActISO/IEC 42001NIST AI RMFGDPR and data residencyAudit records

FAQ

Questions we get asked before starting

Which of the six do we need?

Usually two or three together. Retrieval plus testing is the most common starting pair. Private is added when a rule requires it, and custom training is the last one to reach for rather than the first.

Can you take over something already built?

Yes, and it is a large part of this work. We start by measuring what is there, which often finds that the model is fine and the retrieval or the integration is not.

Do you use one model provider?

No, and we build so that you do not either. Everything talks to a single internal endpoint, so changing model or provider is a configuration change rather than a rebuild. That is a deliberate design decision from the first week.

How does this relate to your workflow automation work?

Same practice, different problem. Workflow automation is about work your team repeats. This is about building the AI itself, into your systems or your product. Some projects need both and the two sides use the same engineers.

Next step

Tell us what you tried and where it stopped.

If there is a demonstration that never shipped, the reason it stopped tells us more than the demonstration does. If there is nothing yet, tell us what you want it to do and who would rely on it.

Ahmedabad, India. We work with teams in the US, UK, Europe, Singapore and the Gulf, and we are used to the time difference.