Reports that took months now take days

DOCUMENT INTELLIGENCE · RAG PIPELINE · IN PRODUCTION FOR 2+ YEARS

40%

reduction in report generation time

> 0.83

BERTScore on evaluation data

100%

of answers cite their source documents

2+ yrs

in production and daily use

Client: HecoAnalytics, health economics & market access
Users: Health economists producing client reports
Solution: AI document search + reporting
Stack: Python · LLMs · RAG · AWS
My role: Data Scientist: design, build, evaluation and deployment
Status: In production

The challenge

Health economists were spending hours searching through research papers, reports and studies to find the information needed for client reports.

And the firm could not rely solely on AI and wait to catch hallucinated results: every answer had to come from a peer-reviewed journal.

The information already existed. Finding, checking and restating it was what consumed the time.

A standard AI chatbot wasn't enough. In healthcare analytics, answers needed to be accurate, traceable and based on the firm's own research.

The solution

I built an AI system that searches the company's research documents and uses the relevant evidence to generate answers.

Every answer includes its source, so the economist can quickly check where the information came from.

Before deployment, I tested different models and configurations against real questions from the team to measure accuracy.

The system was then deployed into the client's reporting workflow.

PythonRAGLLM evaluationBERTScore / ROUGEAWSVector search

The results

40% faster report generation

The system handles much of the time-consuming work of finding and extracting evidence, allowing economists to spend more time on analysis and judgement.

Trusted and used in production

The system has been running for 2+ years and is used across client-facing workflows.

Evidence-backed AI

Answers are linked back to their source documents, making the output easier to verify and trust.

The key wasn't simply adding AI. It was making the answers verifiable.

"Abrar is a skilled programmer and data scientist who works well as part of a team. Abrar acquires knowledge effectively and always demonstrates excellent work across a number of platforms (including cloud)."

DAVID BELL · CTO, HECOANALYTICS · READER IN COMPUTER SCIENCE, BRUNEL UNIVERSITY OF LONDON

Why it worked

Source citations. Users can verify the information.
Human in the loop. AI enhances the health economists' work, but they stay in the driving seat.
Measured before deployment. Accuracy was tested on real client tasks.
Built around existing documents. No need to replace the team's knowledge base.
Expanded over time. The same foundations were later used for summarisation and other research tools.

Could this work for your business?
If your team spends hours searching through documents, reports, policies or research for information you already have, there may be a simpler way. I'll tell you honestly whether AI can help, or whether the documents need fixing first.

Next case study: an automated content engine, 10K followers in 3 months →