January 2026 – August 2026
Jan 2026 – Aug 2026
Coming back for a second stint meant owning what we shipped — not starting fresh.
Cost Optimisation — ~10x token reduction Migrated the AI performance dashboard from GPT-4.1 to Gemini 2.5 Flash. The hard part wasn't the model swap — it was re-engineering prompts from scratch across vendors while keeping outputs indistinguishable. Result: a fraction of the original inference cost at the same quality bar.
Latency POC Prototyping a restructured workflow for the enterprise chatbot to reduce response latency. Currently in progress.
BAU Day-to-day ownership of two production AI systems serving 5,200+ financial consultants — monitoring, bug fixes, edge case handling.
Owning production systems is different from building them. Every decision has downstream consequences at scale. That accountability sharpens how you think.
May 2025 – August 2025
I joined Prudential twice. The first time, I built things from scratch. The second time, I owned what we shipped.
Between May 2025 and August 2026, I spent 12 months inside Prudential's AI team — not as a peripheral contributor, but as the person who designed, built, stress-tested, and maintained two production AI systems that 5,200+ financial consultants use daily.
Financial consultants at Prudential deal with an enormous product catalogue — policies, riders, terms, exclusions, benefit tables. When a client asks a nuanced question, consultants need the right answer fast. The old way: search through documentation, call a colleague, or escalate. None of that scales across thousands of advisors.
The AI Chatbot (WhatsApp)
We designed and shipped an enterprise RAG + MCP-powered chatbot that lives inside WhatsApp — where consultants already work. It retrieves product and policy knowledge on demand, 24/7, without a human on the other end.
The retrieval pipeline uses Azure AI Search with vector embeddings. I built the knowledge base ingestion pipeline, chunking strategy, and retrieval configuration. I also ran the full testing suite — unit, edge-case, and stress tests — before we pushed to production in June 2025.
The AI Performance Dashboard
Alongside the chatbot, I co-built an AI-powered performance dashboard that analyses each consultant's activity data and generates personalised insights — specifically, their shortfall against the next award tier and actionable suggestions to close the gap.
This one ran on a batch-processing architecture: data collected overnight, AI-generated insights ready by morning. It went live in September 2025.
| Layer | Technology |
|---|---|
| Retrieval | Azure AI Search, vector embeddings |
| LLM | GPT-4.1 |
| Orchestration | MCP, RAG pipeline |
| Channel | WhatsApp Business API |
| Testing | Unit, edge-case, stress |
Coming back for a second stint meant I wasn't starting over — I was responsible for keeping two production systems alive and making them better. That's a different kind of pressure.
The AI performance dashboard was running on GPT-4.1. Good model, expensive at scale. My task: migrate to Gemini 2.5 Flash without degrading output quality.
The hard part wasn't swapping the model. It was that GPT and Gemini are different vendors with different prompt semantics — what works for one doesn't translate directly to the other. I had to re-engineer the prompts from scratch, run systematic output comparisons, and iterate until the results were indistinguishable from the original.
Result: ~10x reduction in token costs. The dashboard now runs at a fraction of its original inference cost.
I also prototyped a latency optimisation architecture for the chatbot — restructuring the workflow to reduce response time. This is currently a POC and not yet in production, but the groundwork is laid.
Beyond the headline work, I own the day-to-day: monitoring, bug fixes, edge case handling, and keeping both systems stable across the consultant population.
Building for 5,200 users is different from building for yourself. Every design decision has downstream consequences at scale — a bad chunking strategy means wrong answers for thousands of consultants. A prompt that degrades on migration means inaccurate performance insights for thousands of advisors.
It forced me to think in systems, not features. And it gave me a reference point for what "production AI" actually means — not a demo, not a prototype, but something people depend on every day.