Case Studies
Real projects, real numbers
Six production AI builds across sports, education, enterprise research, sales operations, event technology, and maritime procurement. Client names are anonymised on the site, but the systems, the architecture, and the results are real. Every project is still running.
Get new case studies as we publish them
One email per published case study. No marketing drip.
Building a Bloomberg-Style AI Sports Prediction Platform
Sports Prediction Platform · Sports Technology
Sole engineer on a full-stack AI prediction platform for NBA and PGA golf, combining real-time data feeds, machine learning models, prediction market analysis, and a conversational AI layer.
The Challenge
The client wanted a Bloomberg-style sports intelligence platform: real-time data, predictive models, market edge calculations, and actionable insights for subscribers. Nothing existed, it needed to be built from scratch, fast, and to commercial quality.
Our Approach
I took full technical ownership as the sole engineer. Designed the data architecture, built the ML pipeline, created the frontend dashboards, and deployed the entire system. The key architectural decision was separating the prediction engine (Python/FastAPI) from the presentation layer (React/TypeScript) to allow independent scaling.
What We Built
Built end-to-end: FastAPI backend with SQLAlchemy ORM, React frontend with real-time dashboards, Elo-based backtesting engine, market edge calculations using Kelly criterion, Polymarket integration for prediction market data, and Claude AI integration for natural language queries.
Results
- 100+ daily predictions generated automatically
- NBA MVP prediction model and PGA golf expansion launched
- Live commercial product with paying subscribers
- Full-stack platform designed, built, and deployed in under 3 months
- Claude AI chat integration for conversational sports intelligence
Scaling Culturally Accurate AI Image Generation for Education
Education Technology Platform · Education Technology
Designed and built an AI image generation pipeline that produces thousands of culturally appropriate educational illustrations for an African language learning platform, at a fraction of the cost of manual illustration.
The Challenge
The client needed 2,400+ educational images across 14+ categories, animals, food, household items, cultural scenes, body parts, all culturally appropriate for African language education. Manual illustration was prohibitively slow and expensive.
Our Approach
The core challenge was not image generation, it was cultural accuracy at scale. I designed a 5-tier prompt engineering system where each category has specific rules. Smart reference selection ensures style consistency across thousands of images.
What We Built
Built on Google Apps Script with Gemini 2.0 Flash: 14+ automatic category detection, 5-tier prompt engineering with category-specific cultural rules, queue-based batch processing with fault tolerance, smart reference selection, and automated Google Drive organisation.
Results
- 2,422 total images planned, 646+ completed and delivered
- 14+ automatic content category detection
- 5-tier prompt engineering system for cultural accuracy
- 80% reduction in image production time vs manual illustration
- Fault-tolerant batch processing, resume where you left off
Enterprise AI Evaluation: From Inconsistent to Reliable
Enterprise Research Intelligence Firm · Enterprise AI / Research Intelligence
Built the multi-model LLM pipeline, evaluation framework, and guardrail system that became the backbone of an enterprise AI research company.
The Challenge
the client was generating research insights for enterprise clients using LLMs, but output quality was inconsistent. There was no evaluation methodology, no quality benchmarks, and no way to catch hallucinations.
Our Approach
The fundamental insight was that enterprise AI needs the same rigour as enterprise software: testing, benchmarking, regression detection, and quality gates.
What We Built
Designed a multi-model orchestration pipeline (ChatGPT, Perplexity, Grok) with cross-model verification. Built the evaluation methodology from scratch: golden datasets, regression tests, red-teaming protocols, and structured quality scoring.
Results
- 40% reduction in research delivery time
- 30% improvement in factual reliability of outputs
- Enterprise-grade evaluation framework still in use
- Ongoing engagement since May 2025, now core infrastructure
- Zero hallucination incidents on evaluated outputs
AI-Powered CRM and Growth Engine for SaaS Startup
SaaS Growth Company · SaaS / Sales Operations
Designed and deployed a complete AI-powered CRM and growth infrastructure for an Australian SaaS startup, connecting 6 platforms into a unified pipeline with 16+ live AI automations, 5-path intelligent lead routing, and zero lead loss.
The Challenge
the client was preparing for public launch with a fragmented sales operation: leads entering from multiple sources with no consistent structure, manual follow-ups leading to missed opportunities, no unified pipeline view, and a process that would not scale without hiring more people. They needed a system that could run reliably without constant manual attention while still feeling personal and high quality.
Our Approach
Rather than recommending off-the-shelf workflows, we designed a coherent system architecture with clear stages and decision logic. We started by mapping every lead source and touchpoint, then built a multi-layered automation system with AI embedded at key decision points. Every component was designed to work independently but feed into a single unified pipeline.
What We Built
We built a 6-platform integration (Wix, Zapier, HubSpot, Apollo, Slack, Gmail) with AI agents handling email generation, lead intelligence, and decision support. The system includes 5-path conditional lead routing, 16+ live HubSpot automations, AI-powered personalised email drafting, automated lead summaries, real-time Slack notifications, unified inbound and outbound pipelines, and a comprehensive operational handbook for team handover.
Results
- Zero lead loss across all entry points
- 16+ AI-powered automations deployed live in HubSpot
- 5-path conditional lead routing system
- 6 platforms integrated into unified pipeline
- 3-4x increase in lead handling capacity per person
- AI generates personalised email drafts for every qualifying lead
- Full operational handbook and walkthrough videos delivered
- 6-step cold outreach sequence with automated stop rules
AI Product Content Pipeline for UK Industrial Tooling
UK Industrial Tooling Brand · E-Commerce / Manufacturing
Built a 5-stage AI pipeline that generates marketplace-ready product content and images for 2,000+ SKUs across Amazon UK, eBay UK, and Shopify, replacing manual copywriting and photography.
The Challenge
the client (Woodford Woodworking Tooling Ltd) is a UK manufacturer in Bramhall, Cheshire, selling bandsaw blades, planer blades, and accessories across Amazon UK, eBay UK, and Shopify. With 2,000+ SKUs, creating unique, optimised product listings and professional images for each was taking weeks of manual work per batch. Each marketplace has different requirements for titles, descriptions, and image specs.
Our Approach
We designed a 5-stage AI pipeline: market research (Perplexity), content generation (GPT-4), quality assurance with automated scoring, document output to Google Drive, and AI image generation (Gemini). Each stage runs independently with automated quality gates. The system processes products in batches, respects API rate limits, and includes full error recovery.
What We Built
The pipeline generates store-specific content (Amazon titles, bullets, A+ modules, eBay item specifics, Shopify descriptions with SEO), validates against marketplace rules, auto-fixes issues, and produces 9 professional product images per SKU. Everything outputs to structured Google Docs in organised Drive folders, ready for upload to each marketplace.
Results
- 2,000+ SKUs processed through the pipeline
- 5-stage AI pipeline: Research, Generate, QA, Output, Images
- 9 AI-generated images per product (hero, lifestyle, detail, comparison)
- 3 marketplace formats from single input (Amazon UK, eBay UK, Shopify)
- Automated QA scoring with 8.0/10 minimum threshold
- Machine compliance checking (prevents Amazon listing removal)
- Full Google Drive integration for team review workflow
- Brand voice enforcement with banned word detection
“The structure looks great. This is exactly what we needed to scale our product listings across all our marketplaces.”
Founder
Industrial Tooling Founder
Zero-Hallucination AI for Marine Procurement RFQs
Maritime Services Company · Maritime / Procurement
A maritime services company was drowning in manual procurement. Buyers were spending hours per vendor turning vague keyword briefs (sometimes one word, sometimes a full vendor description) into five structured, vendor-quotable RFQ enquiries, each with real part numbers, specs, and maker/model references. We built a production AI pipeline that does this end to end in under 90 seconds, with zero hallucinations on critical marine equipment data.
The Challenge
Marine procurement is technically unforgiving. A single hallucinated part number, wrong maker, or banned phrase in a quote request damages vendor relationships and costs real money. The team had tried off-the-shelf LLM tools and generic assistants, and every one failed on the specifics: hallucinated Sulzer part numbers on Wartsila engines, ignored safety categories, invented equipment that did not exist. They needed production-grade reliability, not a demo.
Our Approach
Discovery first: we mapped the real workflow a buyer runs end to end, not the idealised one. Then we designed a 10-step pipeline that separates deterministic code from AI calls. Category classification, equipment selection, and catalogue lookups run as database queries against curated, branded data. AI handles only what AI is good at: planning items, writing vendor-quotable descriptions, and validating the result. Every stage has a quality gate, banned-phrase stripping, and a 0-100 score before anything ships to the buyer.
What We Built
A 10-step AI pipeline that turns keywords into five vendor-ready RFQ enquiries per group, each with technically accurate items, real part numbers drawn from curated catalogues, and per-category spec enforcement. Runs on Next.js serverless API routes with Supabase (PostgreSQL + pgvector + RLS), Claude Sonnet for the generative steps, and a Google Apps Script bridge that outputs each enquiry into a formatted Google Doc ready for the buyer to email. Role-based access, full audit trail of AI costs, and a quality-gate score on every run.
Results
- 95%+ pass rate on the internal quality gate across 14 marine procurement categories
- 30-90 second end-to-end processing per RFQ group, down from hours of manual work
- 3,826 equipment entries and 3,397 catalogue items queried live per run for real part numbers
- Zero hallucinated Sulzer/Wartsila/MAN part numbers across thousands of generated enquiries
- Per-run AI cost tracking at roughly $0.15 to $0.30, making unit economics transparent from day one
- Deployed to production on Vercel Pro with 300s function timeouts and full role-based access control
“The speed and reliability changed how our procurement team works. We went from manual RFQ writing that took hours per vendor to fully structured, vendor-quotable enquiries in under two minutes. The zero-hallucination guardrails were the dealbreaker, and the system just nailed it.”
Head of Procurement
Maritime Services Company
Have a project that needs AI expertise?
We bring the same rigour and production-grade engineering to every project we take on.
Start a Conversation