AI systems built for real-world problems

Selected work applying document intelligence, information extraction, retrieval-augmented generation, generative AI and workflow automation to problems teams were solving by hand. Client names are withheld; the engineering is described as it was built.
  • Home
  • /
  • Case studies

Featured case studies

Production systems and product work, each with a fuller write-up covering the challenge, what we built and how it works.

Agentic AI · Governed text-to-SQL · Multi-tenant SaaS

Agentic Procurement Analytics & Approvals

A financial operations platform already held the supplier records, purchase orders, contracts and audit history its customers needed — but every question that was not already a screen became a ticket for an analyst. We built an assistant that writes and validates its own SQL against each tenant’s governed views, answers only from the rows it gets back, and stops for human confirmation before approving or rejecting anything.

  • LangGraph
  • Azure OpenAI
  • Microsoft SQL Server
  • FastAPI
Read case study
  • In productionacross multiple enterprise tenants, each sealed to its own governed views
  • Confirmation gatebefore any approval, rejection or state change is written
Document intelligence · RAG

Legal Document Intelligence Platform

Litigation teams at a law firm were spending around 60% of their time organising documents rather than building legal strategy, with manual review and timeline construction taking two to three weeks per case. We built a system that extracts the parties and maps their relationships as a knowledge graph, generates case timelines with automated dispute detection, drafts documents from the case material, and answers questions about the case conversationally.

  • Gemini
  • RAG
  • AWS Textract
  • React
Read case study
  • 85%reduction in document review time, from 2–3 weeks to 2–3 days
  • 90%improvement in timeline accuracy with automated dispute detection
RAG & knowledge systems

ShamelaGPT

Researchers in Islamic studies had no efficient way to verify claims against the Shamela digital library, one of the largest repositories of Islamic texts. ShamelaGPT fact-checks: submit a claim, including as an image, and it cross-references the corpus, returns a verdict grounded in the documents it found, and links every source so the reader can verify it independently — on top of semantic, multilingual search across the whole library.

  • RAG
  • Semantic search
  • Multilingual
Read case study
  • Fact-checkingcross-referenced across books, with links to every source
  • Livein production at shamelagpt.com
Climate risk analytics · RAG · CSRD/ESRS reporting

Climate Risk & Adaptation Intelligence

A property owner sees a flood risk score of 68 out of 100 — accurate, and unusable on its own. We built an advisor that sits between live climate-risk data and a catalogue of 97 adaptation products across five hazard classes, holding a conversation instead of serving a long questionnaire. Any conversation exports as a CSRD/ESRS E1 climate risk and adaptation report.

  • FastAPI
  • PostgreSQL with pgvector
  • Gemini 2.5 Flash
  • Server-Sent Events
Read case study
  • 0.81stime to first byte, six times faster than the naive pipeline
  • 97adaptation products matched across five hazard classes
Document intelligence · Information extraction

Enterprise Invoice & Receipt Processing

Enterprise invoice workflows required three to five days of review per document, with operational cost, bottlenecks and inconsistent extraction accuracy across differing document formats and languages. We built a multi-model extraction pipeline — layout-aware form recognition, custom named-entity recognition, and an LLM pass for validation and enrichment — handling PDFs, images and scanned documents with multi-language support and multi-tenant isolation.

  • Azure Form Recognizer
  • GPT-4
  • Custom NER models
  • Multi-language
Read case study
  • 90%+reduction in processing time, from 3–5 days to minutes
  • 95%+field extraction accuracy through multi-model validation
  • 99.8%processing reliability with comprehensive error handling
Automation · Document intelligence

Multi-Supplier Invoice Processing Automation

MPI Plumbing Corp was processing thousands of invoices a day across more than twelve different supplier formats, with manual PDF-to-Excel conversion and reconciliation taking three to five days per batch. We built an end-to-end pipeline that collects invoices straight from an inbox, identifies the supplier, maps line items to standardised product codes, handles tax distribution, and posts into the accounting system.

  • FastAPI
  • PostgreSQL
  • Google Gemini
  • QuickBooks integration
Read case study
  • 85%less processing time, from 3–5 days to hours
  • 99.5%extraction accuracy across 12+ supplier formats
Conversational AI · Generative AI

AI Travel Assistant on WhatsApp

Travel planning normally means forms, filters and multiple screens. We built a conversational assistant inside WhatsApp so travellers can describe the trip they want in their own words, then question, refine and change the itinerary in the same conversation — turning planning into an iterative dialogue rather than a one-time generation step.

  • Conversational AI
  • LLMs
  • WhatsApp integration
  • NLP
Read case study
  • Livein production, planning trips inside WhatsApp
  • Iterativerefine and change the itinerary through follow-up messages
Generative AI · Recommendation & personalisation

AI Travel Planning Engine

A useful travel itinerary is not a list of places — it has to account for preferences, available time, destinations, and how places and experiences relate to one another. We built a system that turns a traveller's natural-language requirements into a structured, personalised day-by-day plan, treating itinerary creation as a planning and recommendation problem rather than text generation.

  • Generative AI
  • LLMs
  • Recommendation systems
  • AI planning
Read case study
  • Natural languagerequirements in, structured day-by-day plan out
  • Personalisedrecommendations, with a foundation for learning from behaviour
Computer vision · On-device inference · Privacy

On-Device Bystander Blurring for Live Video

Shop staff hold live video calls with customers from the shop floor, and everyone else in frame is a bystander who never agreed to be on camera. We built a model that recognises enrolled staff — from a short video the shop owner uploads — and blurs everyone else during the call, running on the device, with models distributed per store.

  • Computer vision
  • On-device inference
  • Android
Read case study
  • On-deviceblurring during the live call, not in post-processing
  • Per-storemodels pushed by store and staff identity

Further work

Additional production systems. Detailed write-ups for these are not yet published.

Document intelligence · Banking

Multilingual Purchase Order Extraction

An information extraction system for scanned purchase orders in a banking environment, handling both Latin and Cyrillic scripts and pulling key fields from scans automatically, replacing manual data entry.

  • OCR
  • Multilingual
  • Cyrillic & Latin scripts
AI agents · Analytics

Supply Chain Analytics Agent

A conversational agent over large volumes of supply chain invoice data spanning multiple companies. Natural-language questions return analysis of spending patterns, vendor performance and financial trends across the network, and generate analysis reports automatically.

  • Conversational AI
  • Data analytics
  • Reporting
Automation · RPA

RPA Data-Entry Automation

A robotic process automation workflow replacing manual data entry in business process outsourcing operations. It extracts data from varied sources, validates it, and populates multiple downstream systems and databases.

  • UiPath
  • RPA
  • Validation

Have a problem like one of these?

Tell us what you are trying to automate, extract, analyse or build, and we can work out what an appropriate system would look like — and what it would cost.

Start a conversation See packaged options