Eeshan GuptaAI Engineer — GenAI, RAG & Voice AI Systems

Building an AI product and shipping it is easy. Knowing how and why it works is the real challenge. If you don’t know how or why you did something, you end up facing problems you never knew could exist.

I build the parts of AI products that have to hold up in production — retrieval pipelines, LLM evaluation harnesses, and voice/ASR systems — for pharma, legal, logistics, and customer-support teams at enterprise scale.

Project stories

Where the numbers came from

Each figure below comes from a project I worked on.

5x

Faster authoring on a GenAI document-drafting platform for a top pharmaceutical client. No technical skills were needed to operate it.

Problem
Authoring was completely manual. Each document covered analytics for a marketed drug or experiments for one still in development, and its parts were owned by different teams: legal, sales and marketing, and R&D. The back and forth between them meant even a first draft took six months, while competition made getting the document to the authorities urgent.
Constraint
Each part needed information from other parts of the same document, and sometimes from other documents. So generation worked on the parts of one section at a time, never a whole section in one go.
Decision
We started where cross-team collaboration was heaviest and built a way to generate a first usable draft of those parts within a few hours. Working one section at a time also meant its parts could be generated in parallel.
Result
5x faster authoring.
90%

A large-scale NLP extraction pipeline for a top pharmaceutical client that reduced manual codification effort.

Problem
A third-party service provider codified informed consent forms (ICFs) by hand for compliance. Even within one study, every site could have its own ICF, and each form had several items that had to be extracted and codified manually. That meant days of work for a single study and a large bill from the third party.
Constraint
It had to be as quick, as accurate and as cheap as possible, and a person always approves the codification.
Decision
We ran an orchestrator for every study, which in turn runs multiple workers that extract each site’s items in parallel.
Result
Manual codification effort cut by 90%.
Up to 50%

A RAG-based customer support system for a leading global travel technology provider.

Problem
Airline staff use the provider’s suite of software, and it takes training to use. Even after training, staff still had questions, which they raised as tickets. A support executive answered each one, usually by looking it up in the user manuals.
Constraint
Answers had to be accurate and quick. Most tickets were about a single product.
Decision
We built a chatbot into the product interface to answer staff in place of the support executive, and only a few queries still became tickets. A standard RAG setup behind it didn’t give good answers, so we did a root cause analysis. Different prompts and better models helped only slightly. Then we looked at ingestion and found that retrieval was returning chunks that were only partly correct. Once we used hierarchical chunking, the answers became more accurate. They were also more complete, which had been missing before.
Result
Support ticket volume cut by up to 50%.

Open questions

What I'm working through

What does RAG look like beyond a chatbot?

RAG is mostly used behind a chatbot, but it has many other applications. And no two RAG systems are built the same: the ingestion pipeline changes, and so does the retrieval.

How do you make subjective LLM evals quantitative and understandable?

Evaluations tell us how an ML or deep learning algorithm is performing. LLMs, and products built on them, are less deterministic, so they need evaluations too: formulaic ones, and subjective ones as well. Making those subjective evaluations quantitative and understandable for the end user is the problem I’m interested in.

Get in touch

Write to me if…

  • you’re using RAG for more than a chatbot, or want to talk through how your ingestion and retrieval are built.
  • you need evaluations for an LLM product, including subjective ones that are quantitative and easy for end users to understand.

Outside the day job

Projects and collaborations

I occasionally take on a small, focused project, usually evaluating or hardening a RAG or voice AI system.

LLM & voice model evaluation

Comparing LLMs, ASR, and RAG pipelines on real data: accuracy, latency, and cost, and a clear answer on which one holds up.

Compare notes

Hardening a private RAG system

Single-tenant retrieval running inside a company's own AWS or Azure environment, with role-based access and audit logging from the start.

Compare notes