AI projects we have delivered

Problem type

Personalisation at scale

Service area

Recommendation Systems

Used by

Travellers planning trips

The challenge

Travellers wanted destination suggestions that matched who they are — adventure, business, nature, family or couples — not a generic list. It also had to stay fast and affordable as usage grew.

What we built

An end-to-end AI pipeline: data prepared with AWS Glue and Spark, BERT classifiers that profile traveller personas, and multi-agent orchestration on AWS Bedrock combining graph and document retrieval so every suggestion is grounded in real data.

Personalised recommendations in production, with cost and latency cut at scale through prompt caching, model right-sizing and batch processing.

AI approach

Multi-agent orchestrationGraph + vector RAGPersona classificationPrompt caching & model right-sizing

Tech stack

AWS BedrockAWS GlueApache SparkBERT

Problem type

Fragmented multi-role workflow

Service area

Web Portals & Apps

Used by

Administrators, charity directors and donors

The challenge

Admins, charity directors and donors managed donor-advised fund accounts and grant decisions across separate tools, with no shared place to discuss, vote and decide.

What we built

A role-based portal with dedicated Admin, Charity Director and Donor views, built-in chat, an insights view, a voting workflow for funding decisions and AI-assisted lookups — integrated with the custodial platform that holds the fund accounts.

Live and running in production for all three user roles.

AI approach

Role-based accessVoting workflowLLM chat & AI lookupsCustodian platform integration

Tech stack

Web portalLLM chatThird-party API integration

Problem type

Customer & market insight

Service area

Insights Engine

Used by

Optometry equipment manufacturers

The challenge

Optometry equipment manufacturers wanted to understand how their products are received across different demographic groups — evidence sales figures alone could not give them.

What we built

A production chat and insights platform over an enriched knowledge base. Retrieval keeps every answer grounded in source data, and sentiment analysis rolls up into brand-sentiment reports and dashboards.

Manufacturers get grounded, demographic-level insight into how their products are accepted.

AI approach

RAG over an enriched knowledge baseSentiment analysisGrounded report generation

Tech stack

LLM chatVector retrievalInsight dashboards

Problem type

Trend & sentiment monitoring

Service area

Insights Engine

Used by

Product and marketing teams

The challenge

The team needed to know which ophthalmology products, brands and companies were trending in social-media conversations — and how people felt about each of them.

What we built

A pipeline that collects posts from relevant social-media communities, scores sentiment towards each product, brand and company, and turns the results into word clouds and charts.

Detailed trend and sentiment insight on ophthalmology products and brands.

AI approach

NLPEntity-level sentiment analysisTrend tracking

Tech stack

Python NLPWord cloudsCharting

Problem type

Manual document processing

Service area

Document & Content Automation

Used by

Legal staff preparing deed orders

The challenge

Preparing deed orders from title commitment documents was slow and manual, and every deed type needed a different template.

What we built

GPT-4 reads each uploaded title commitment and extracts the grantors, grantees, property address, state and county. A staff member confirms the details, then the correct deed template is filled automatically, emailed through a Make (Integromat) workflow and filed in Google Drive.

10× less manual effort on deed preparation, with 100% precision through human-in-the-loop validation.

AI approach

LLM document extractionHuman-in-the-loop validationTemplate-based generationWorkflow automation

Tech stack

OpenAI GPT-4Make (Integromat)Google Drive

Problem type

Scaling scarce expert knowledge

Service area

Conversational Agents

Used by

Patients

The challenge

The client wanted a visual expert that could talk with patients, help identify their issues and respond with empathy — answering strictly from approved Chinese medical texts, with no made-up advice.

What we built

Approved medical texts in a vector database with RAG on Azure OpenAI; GPT-4 fine-tuned for an empathetic, doctor-style manner; speech-to-text and text-to-speech for voice; and a Wav2Lip lip-synced video avatar served through a FastAPI web app.

Pilot delivered with 100% precision on generated answers and no hallucinations; validated for production development.

AI approach

RAG on approved sourcesLLM fine-tuningSpeech-to-text & text-to-speechLip-synced video avatar

Tech stack

Azure OpenAI GPT-4Vector databaseWav2LipFastAPI

Problem type

Manual prospecting & outreach

Service area

Document & Content Automation

Used by

Sales and marketing team

The challenge

Finding B2B leads and writing tailored outreach took hours of manual research per prospect, which capped how many targets the team could reach.

What we built

An engine that crawls market intelligence on target companies and generates targeted outreach content automatically, ready for the team to review and send.

B2B lead generation and marketing running at scale — designed and built from the ground up.

AI approach

Web crawlingMarket-intelligence extractionLLM content generation

Tech stack

Web crawlersLLM content generation

Problem type

Accurate self-service answers

Service area

Conversational Agents

Used by

Customers

The challenge

Customers needed accurate answers drawn only from the company’s own content — not generic answers from the open web.

What we built

A retrieval-augmented generation (RAG) framework using a vector database and OpenAI GPT-4 that answers only from a curated set of uploaded content.

Customers get answers grounded in internal data only.

AI approach

RAGAnswers restricted to approved content

Tech stack

OpenAI GPT-4Vector database

Problem type

Contextual recommendations & ads

Service area

Recommendation Systems

Used by

Chat-platform users and advertisers

The challenge

An existing chat engine needed to recommend relevant content and ads based on what users were actually talking about.

What we built

An add-on that analyses conversation context with sentiment analysis, named-entity recognition and text categorisation, then recommends content such as movies and books and serves suggestive ads.

Context-based content recommendations and advertising driven by chat history.

AI approach

Sentiment analysisNamed-entity recognitionText categorisation

Tech stack

NLP pipeline

Problem type

Round-the-clock patient support

Service area

Conversational Agents

Used by

Psychotherapy patients between sessions

The challenge

Psychotherapy patients needed empathetic support between sessions, at any hour, when their therapist is not available.

What we built

The open-source Falcon LLM fine-tuned on psychotherapist conversations to create an empathetic chatbot companion.

Psychotherapists can support patients 24/7 with an empathetic AI companion.

AI approach

Open-source LLM fine-tuningEmpathetic dialogue design

Tech stack

Falcon (open-source LLM)

Problem type

Risk prediction & prioritisation

Service area

Insights Engine

Used by

Safety and risk teams

The challenge

Workers’ compensation claims were costly, and there was no consistent way to tell which workplaces were most likely to have safety issues.

What we built

A machine-learning model that scores each workplace’s safety propensity from factory size, employee count, machinery size and weight, and audit history — with results delivered in Power BI dashboards.

Workplaces ranked by safety risk, so interventions go where they matter most.

AI approach

Predictive modellingRisk scoringBI dashboards

Tech stack

Pythonscikit-learnAWSPower BI

Problem type

Up-to-date domain guidance

Service area

Conversational Agents

Used by

Viewers and retail crypto users

The challenge

Viewers wanted to understand crypto, with answers and suggestions that keep pace with a fast-moving market.

What we built

A chatbot trained on web data and crypto domain knowledge that gives users insights and suggestions in plain language.

AI approach

LLM chatbotWeb data ingestion

Tech stack

LLMWeb data pipeline

Problem type

Guided product discovery

Service area

Recommendation Systems

Used by

Online shoppers

The challenge

Shoppers needed a guided, conversational way to find the right products.

What we built

A chat application that recommends products through natural conversation.

AI approach

Conversational AIProduct recommendation

Problem type

Local discovery

Service area

Conversational Agents

Used by

Local customers

The challenge

Customers needed a quick way to discover local events and relevant products.

What we built

A chatbot assistant that recommends local events and products in conversation.

AI approach

Conversational AIRecommendations