IntellectAI · 2022–2024

First production LLM in regulated finance

Proved solo. Led a team of eight to production.

Product Strategist & LLM Lead · 2022–2024

In September 2022, I wrote the tender that won IntellectAI the ESG data automation contract for the world's largest pension fund, $2.2Tn in AUM, 9,000 portfolio companies. Four competitors entered the competitive dialogue. We won on quality and risk, which together made up 75% of the award criteria.

The brief: build a system to extract ESG intelligence across the fund's entire portfolio from publicly available documents. When I got into what that actually meant, the problem was bigger than the brief suggested.

External ratings providers covered less than 10% of the portfolio on topics like water and biodiversity. Data arrived 12 to 18 months stale. The fund's analysts could cover only 60 to 80 portfolio companies in depth each year. At 9,000 companies, that left the vast majority either lightly skimmed or not reviewed at all; full coverage at that level of rigor would have required over 100,000 analyst days annually, a scale no team could sustain.

That wasn't a data gap. It was a broken paradigm.

Code red

Eleven months in, the ML pipeline was failing. Accuracy fell below what the client needed. The contract was at risk.

I proved an LLM-based replacement myself and benchmarked it on the same accuracy criteria the ML approach had been measured against. The results held up. I escalated to the CTO, who gave me a team of eight. The ML team had spent 18 months on this. Eight of us shipped the replacement in weeks.

On 15 August 2023, the CTO sent this to the full organisation. Subject: "Thank you and Congratulations for delivering the first Production outcome using LLMs." Probably the first in regulated finance anywhere.

What it unlocked

The LLM QA framework I built during the rescue became the retrieval foundation for what is now the Enterprise Knowledge Garden, the data layer Purple Fabric runs on. The pipeline scales from a single company URL to a fully searchable knowledge base in four stages: crawl, preprocess, classify, embed.

In January 2024, four months before Purple Fabric's official launch, I ran Intellect's first Prompt-a-thon: three days, 20+ ESG analysts learning to build with LLMs from scratch.

Four months after that, the fund's lead analyst sent the first formal commendation we had received in 18 months of service: "Almost all of the questions meet or exceed our expectations. We are already working on putting 12 of 13 questions into production."

That was the first time they put any Intellect-delivered data into production.

The outcome

The system processes queries against 60 billion+ searchable chunks from 10 million+ documents, with full source lineage to the original filing. All 9,000 portfolio companies, covered in depth. The work is equivalent to 100,000+ analyst days annually, a scale no human team could sustain at reasonable cost, accuracy, or speed.

The team ran accuracy checks against SBTi, CHRB, Climate 100, and Net Zero Tracker. The system matched or exceeded all four. Most discrepancies turned out to be data the benchmarks had missed, not errors in the output.

The ML team had 18 months and a 20-person organisation. Eight people, a few weeks. What began as a contract rescue became the blueprint for how Purple Fabric approaches knowledge retrieval today.