AI features your product needs — built to still work the day after the demo.

Yes. A significant share of our AI work is built for other agencies and consultancies under NDA. Three partnership models: white-label (your brand, our engineers, fully invisible), agency-of-record (a dedicated AI pod working as your in-house capacity), and capacity overflow (sprint-by-sprint engagement when your team is full). Code ownership transfers to your repositories, and we run inside your tooling — Slack, Jira, Linear, ClickUp, Asana — as standard. AI is the work agencies are most often asked for and least often staffed for, and this is how we fill that gap.

Our entire team is based in Noida, India — 40 people in our iThum Tower B office, founded in 2015. We work with product teams and agencies across the UK, US, Ireland, Australia, the UAE, Germany, and the Netherlands. Working hours overlap with UK mornings, the full EU workday, and US afternoons. For agency partners we run in their tooling as standard, and every engagement has a named project lead as a single point of contact.

For most AI integrations, RAG — retrieval-augmented generation. RAG gives the model the right information from your content at the moment it answers, handles knowledge that changes, and can cite its sources. Fine-tuning changes how a model behaves rather than what it knows, and it does not reliably teach new facts. We reach for RAG first, prompt engineering before that, and fine-tuning only when objective evaluation proves it is needed. We wrote a full piece on the decision in our journal.

AI integrations range from a single focused feature through to a product-wide AI capability with several integrated features. The scope drivers are the number of features, the messiness of the data the AI must work from, the accuracy and latency targets, and the eval coverage required. We scope every engagement against the specific brief, are competitive with established engineering rates internationally, and are honest about which features should ship in phase one and which can wait. An AI-readiness audit is the lowest-commitment way to get a costed roadmap.

A focused RAG feature or in-product copilot is typically an 8 to 14 week engagement, depending on how clean the underlying data is and how high the accuracy bar needs to be. An AI-readiness audit is 2 to 4 weeks. A semantic search upgrade is often 6 to 10 weeks. We work in two-week sprints with weekly demos, a preview environment, and — because the eval harness is built first — a visible accuracy number the whole way through.

With grounding, verification, and a named ceiling. Every RAG feature answers from retrieved passages, with an explicit instruction to use only that context and to say it does not know when the context falls short. A judge model checks that each claim in an answer is supported by its sources. The hallucination rate is measured against a golden set and held under a ceiling named in the contract — typically around 1%, often well under. We do not promise zero wrong answers; we promise a measured, low, and stable rate.

Evals are how you measure whether an AI feature is actually right. An eval suite is a golden dataset of real questions with known-good answers, plus a judge model that scores every output and a CI gate that fails the build when accuracy drops below a floor. Evals are the single thing that separates AI you can improve from AI you can only hope about — and they are the step most AI work skips because they take real engineering effort. We build the eval harness before the feature, not after.

Whichever the eval data favours for your workload. We build with Claude, GPT-4o, Gemini, and open-source models such as Llama and Mistral, behind an inference gateway that can route between them. We are deliberately not locked to one provider: cheap requests go to a cheaper model, the full path goes to the strongest, and we A/B new releases quarterly because the frontier moves every few months. The model is a swappable component; the engineering around it is what we build to last.

Yes — that is the normal case, and it is the whole idea of AI integration. The AI feature is usually built as a separate service that your existing product calls over an API, so your product is not rewritten; it gains an endpoint or two. The feature ships behind a flag, rolls out to a small audience first, and is A/B tested against the non-AI path. Your product stays your product, and AI becomes one capability inside it rather than a reason to start over.

Your data stays yours. The major model providers offer API terms under which prompts and outputs are not used to train their models, and we build on those terms. Your content for retrieval lives in your own database — pgvector on your PostgreSQL, or a vector database in your environment — not in a model’s weights. Where data sensitivity requires it, we can design around open-source models you host yourself. We agree the data-handling approach explicitly at the start of every engagement.

Yes, and AI features need it more than most software. We offer ongoing engagements covering eval-set growth from real production failures, judge-model evaluation, quarterly model A/Bs and swaps, cost optimisation, and per-feature dashboards for latency, cost, accuracy, and hallucination rate. Model providers ship cheaper and better models every quarter; without a maintenance discipline you drift on cost and miss accuracy gains. An AI feature is a system that needs tending, not a build you walk away from.

AI Integration — this page — is adding AI features to an existing product: RAG, search, copilots, classification. AI Workflow is about automating a process with AI, including multi-step agents that take actions across tools. AI Product is building a product where AI is the core, not an add-on. They share the same engineering foundation — retrieval, evals, observability — and many engagements touch more than one. Our AI Engineering hub page is the place that ties all three together.

Then we will tell you, early, before you have spent a budget on it. Not every problem is an AI problem — sometimes better search, a clearer interface, or a small piece of conventional engineering solves it faster, cheaper, and more reliably. An AI-readiness audit exists partly to catch exactly this. We would rather lose a project than ship an AI feature that should not have been built; an honest “no” is worth more to your business than an impressive demo that does not pay back.