An AI product suite that saved £1.5M and lifted accuracy 30%

The platform held large volumes of regulated safety documentation. Search was keyword-only, document review was manual, and a key-term intake step depended on extraction accuracy that had plateaued in the mid-60s.

Each of these was a direct drain on expert analyst time, the most expensive resource in the business, and a ceiling on how much the platform could scale per customer.

• Quantified the analyst time lost to slow search, manual summarisation and low-accuracy extraction.
• Isolated three distinct AI surfaces, search, summarisation and intake, rather than one monolithic AI feature.
• Defined success up front with an HHH evaluation framework, helpful, honest and harmless, with concrete accuracy targets.

• Ship three targeted AI products on AWS, each owning one bottleneck: AI-Semantic Search, AI-Summarizer and AI-Intake.
• Match the architecture to the job: vector search for meaning, RAG for citable summaries, an NLP microservice for extraction.
• Treat extraction accuracy as a first-class product metric, tracked release over release.

• Built AI-Semantic Search on an embeddings and vector-search architecture, answering meaning-based queries instead of keyword matches.
• Built AI-Summarizer on a RAG architecture, condensing 5,000 to 10,000 word documents to under 500 words with citable references, essential in a regulated domain.
• Built AI-Intake as an NLP entity-extraction microservice, iterated across v1 and v2 with targeted data training.
• Ran every model against the HHH framework before release.

• AI-Semantic Search reduced search time by 90% through meaning-based queries.
• AI-Intake lifted extraction accuracy from the mid-60s to the upper-90s, around a 30% gain.
• AI-Summarizer delivered cited summaries under 500 words, trusted enough for regulated review.
• Together, the three products saved over £1.5M across the suite.

• Define the evaluation framework before building the model, not after.
• Accuracy is a product decision, not only an engineering one.
• In regulated domains, citations are the feature that makes AI usable.