AI & machine learning developmentFeatures that have to be
right, not impressive.

Retrieval, LLM features, classification and forecasting, each with an evaluation harness so you know when it got better and when it did not.

Book a consult See our work

Eval harness first

A number before a demo.

Retrieval done properly

Chunking, ranking, citations.

Cost and latency

Budgeted, then measured.

Engagement

Data & AI · project or retained

Every version is scored on the same questions. It ships when the numbers clear the rule, not when the demo looks good. Example figures.

What you get.

An evaluation harness before a demo: a number you can defend to your board

Retrieval done properly: chunking, ranking, citations, and the cases where it should say nothing

Cost and latency budgeted up front, then measured in production

Monitoring, retraining and rollout owned end to end, not handed to you as a notebook

Engagement: Data & AI · project or retained · Ask about your AI feature →

Good for

— An AI feature that has to be right, not impressive

— Search or support that should answer from your data

— Forecasts the finance team will actually use

We will say no when

— The goal is an AI feature rather than a problem the feature solves

— There is no data you are allowed to use, and no plan to get some

Stack · only what we have shipped to production

Claude CodeMCP serversLangGraphLangChainCrewAIRAGpgvectorQdrantClaude APIOpenAIGeminiAWS BedrockLangSmithLangfusePromptfooTensorFlowOpenCV

Where we do this mostFintech & payments →Retail & e-commerce →

From the record.

All case notes →

01 · Retail · AI

Data & AI

Personalisation and image-search engine

Problem — Recommendations and visual search that had to prove their return in the client’s numbers, not in a demo.

What we did — Recommendation and image-search models with evaluation before release, production monitoring, retraining and version rollout owned end to end.

Measured — Up to 261× client ROI. Shipped an image-search model into production.

Recommendation systemsImage searchEvaluation gatesModel monitoring

Client withheld · details on the callAsk about this →

02 · E-commerce · AI

Data & AI

LLM and RAG assistant in a live customer journey

Problem — A UK deals platform had inbound support volume it could see in its tickets.

What we did — An assistant sourced from support-ticket analysis rather than a feature request, embedded in the live journey, with quality gates before release.

Measured — Reduced inbound support volume — [ADD THE NUMBER].

LLMRAGQuality gatesSupport analytics

Client withheld · details on the callAsk about this →

03 · Agritech · computer vision

Data & AI

Detection from user-submitted photographs

Problem — Non-experts photographing in poor conditions; a model that had to act on what they sent.

What we did — Owned the labelled dataset strategy, the evaluation criteria that gated release, production monitoring and the retraining loop as field data arrived; capture guidance and failure handling treated as part of the product. Localised for low-literacy users.

Measured — 98% detection accuracy. Government incubator funding secured on the minimum viable build.

Computer visionDataset strategyEvaluationRetraining loopi18n

Client withheld · details on the callAsk about this →

04 · AI · delivery

Data & AI

Agentic software-delivery framework on Claude Code

Problem — An engineering practice wanted to deliver faster without losing review, tests and documentation along the way.

What we did — An end-to-end pipeline on Claude Code: project planning, automated epic and story population with progress tracking, autonomous story implementation, a dedicated solutions-architect agent, guardrail enforcement, automated code review and test generation, living documentation, and human-in-the-loop escalation pushed to mobile when a decision needs a person.

Measured — Rolled out to 10 engineers and now the practice’s default path for new work. Two Android apps scoped at five developers were shipped by one, on the original timeline.

Claude CodeMCP serversAgent orchestrationGuardrailsAutomated reviewJira

Client withheld · details on the callAsk about this →

How it runs.

The same loop as every engagement

Week 0

The problem, in a sentence

What the model must get right, how often, and what happens when it is wrong.

Week 1

Eval set and baseline

A few hundred real cases, labelled with you. The number today, before any model.

Week 3

First model behind a gate

Retrieval or classification in a staging environment, released only when the number moves.

Week 6

In production, watched

Monitoring on quality, cost and latency; a retraining loop; a person in the loop where it matters.

Fortnightly

The review

The eval number, the bill and the latency, before and after. Then the next feature.

Questions we get about this.

Whichever passes the eval at a cost you can afford. We have shipped on Claude, OpenAI, Gemini and Bedrock, and on small models where a large one was too expensive. The harness decides, not the logo.

Sometimes, which is why the harness exists. Retrieval with citations, refusal when the sources are thin, and a measured false-answer rate you can see every fortnight.

Yes. Models run in your cloud account or through a provider with no training on your data, and the retrieval index lives with you.

The other seven.

All eight end in something running

01Custom software→02Team augmentation→03Data engineering & analytics→04AI & machine learninghere05Cloud & DevOps→06Legacy modernisation→07Mobile→08QA & test automation→

AI & machine learning · the call

Tell us what it must get right. We will say how to measure it.

Bring the problem, not a brief. If it is outside what we do well, we will say which of our contacts to call instead.

Book a consult See our work

Or email info@hopsig.com — reply within one working day.

What you get.

An evaluation harness before a demo: a number you can defend to your board

Retrieval done properly: chunking, ranking, citations, and the cases where it should say nothing

Cost and latency budgeted up front, then measured in production

Monitoring, retraining and rollout owned end to end, not handed to you as a notebook

Engagement: Data & AI · project or retained · Ask about your AI feature →

AI & machine learning developmentFeatures that have to beright, not impressive.

What you get.

From the record.

Personalisation and image-search engine

LLM and RAG assistant in a live customer journey

Detection from user-submitted photographs

Agentic software-delivery framework on Claude Code

How it runs.

The problem, in a sentence

Eval set and baseline

First model behind a gate

In production, watched

The review

Questions we get about this.

Which models do you use?+

Will it hallucinate?+

Can you use our data without it leaving our account?+

The other seven.

Tell us what it must get right. We will say how to measure it.

What you get.

From the record.

Personalisation and image-search engine

LLM and RAG assistant in a live customer journey

Detection from user-submitted photographs

Agentic software-delivery framework on Claude Code

How it runs.

The problem, in a sentence

Eval set and baseline

First model behind a gate

In production, watched

The review

Questions we get about this.

Which models do you use?+

Will it hallucinate?+

Can you use our data without it leaving our account?+

The other seven.

AI & machine learning developmentFeatures that have to be
right, not impressive.