---
title: Enterprise GenAI, from sales call to
working POC
subtitle: Fine-tuned LLMs, document search,
multimodal document AI, and the
client conversations that scoped them
slug: enterprise-genai-presales-to-poc
hook: 'Turning "can AI do this?" into scoped,
shipped work: fine-tuned LLMs,
enterprise document search, and multimodal
document AI that cut manual review
time by ~40%.'
role: Senior Data Scientist; GenAI team lead;
pre-sales lead
employer: Royal Cyber Inc. (Lahore)
client: Enterprise clients in North America,
Canada and the Middle East
period: Dec 2022 – May 2024
tracks:
- agentic-genai
- full-stack-fde
- data-science
tags:
- fine-tuning
- LoRA
- PEFT
- multimodal RAG
- document intelligence
- pre-sales
- solution consulting
- team leadership
metrics:
- value: ~40%
label: Less manual review time with
multimodal document intelligence
(text,
vision and structured data)
when: Dec 2022 – May 2024
---
## Context
In 2023 every client asked the same question:
"Can AI do this?" At Royal Cyber I was
often the one answering it. First on
pre-sales calls with clients in North
America, Canada and the Middle East, then
with a proof of concept, then with the
build.
I was a senior data scientist there from
December 2022 to May 2024. Before I moved
on, I also led a team of junior engineers
building GenAI proofs of concept.
## The problem
Most GenAI requests arrive as a wish, not a
spec. "Read our documents." "Answer our
customers." "Summarize the contracts." The
job is to find the part that's real, prove
it fast, and be honest about the rest.
## Constraints
- **Enterprise documents** in legal, financial
and customer-support settings: messy
scans, tables and images, not clean text.
- **Cloud and on-premises deployments**,
depending on the client.
- **The 2023 model landscape:** GPT, LLaMA,
Falcon and Mistral, with limited GPU
budgets. That made parameter-efficient
fine-tuning the practical route.
- **Proofs of concept had to persuade.** A
demo that only works on the happy path
doesn't survive a client's first real
document.
## What I built and did
**Pre-sales.** I took part in, and then led,
pre-sales client calls: listening, finding
the actual problem, proposing a tailored
solution, negotiating, and keeping the
technical plan tied to the client's goals.
**Fine-tuned LLMs.** I fine-tuned GPT, LLaMA,
Falcon and Mistral models with LoRA and
PEFT, then put them to work:
- **enterprise document search** using
dense-retrieval RAG on FAISS and Pinecone;
- **summarization** for legal and financial
documents;
- **customer-support chatbots.**
**Multimodal document intelligence.** Systems
that read documents using text, vision and
structured data together, so people review
only what the model is unsure about.
Manual review time fell by ~40%. The same
extracted content fed multimodal RAG
across text, images and tables.
**A small team, real POCs.** I led a team of
junior engineers building GenAI proofs of
concept for real-time applications, and
explained their value to clients in plain
language.
## How it fits together
Scope first. A pre-sales call turns a wish
into a scoped proposal and a
proof-of-concept plan, and the team builds
the POC.
For document use cases, documents (scans,
PDFs, tables, images) go through
extraction: OCR, layout and tables. Text
and vision models classify each document
and pull out its fields. Anything below a
confidence bar goes to a human; everything
else flows on. Extracted content is
indexed for multimodal retrieval, and
fine-tuned models handle domain language
for summaries and support. The result
shows up as document search, summaries and
chatbots that answer with sources.
## Stack
- **Models:** GPT, LLaMA, Falcon, Mistral;
fine-tuning with LoRA and PEFT; Hugging
Face Transformers; PyTorch, TensorFlow and
Keras.
- **Retrieval:** dense-retrieval RAG and
multimodal RAG; FAISS, Pinecone;
embeddings.
- **Documents:** OCR, layout and table
extraction, vision models.
- **Platform:** Python, Microsoft Azure (Azure
Machine Learning, Azure Functions, Azure
Kubernetes Service), cloud and on-premises
deployment.
## Results
- **~40% less manual review time** with
multimodal document intelligence.
- **Pre-sales calls led** across North
America, Canada and the Middle East.
- **A team of junior engineers led** through
GenAI proofs of concept.
- **Star Performer of the Year 2023** at Royal
Cyber.
- **No recorded figures for the fine-tuning,
search or chatbot work,** so there are
none here.
## What I'd do differently
- **Write the eval set in the first call.**
The fastest way to scope a GenAI project
is to ask for twenty real examples and
what "right" looks like for each. Today
that's the first deliverable, before any
model.
- **Retrieve better before fine-tuning.** In
2023, fine-tuning felt like the serious
option. Most of those wins now come from
better retrieval, structured outputs and a
tight eval loop. Fine-tuning comes after,
if at all.
- **Measure reviewer time from day one.** The
~40% came from reviewing less. Timing
reviewers per document from the first
pilot makes that kind of number airtight.
- **Use the 2026 managed pieces.** For
extraction, I'd now start with Azure AI
Document Intelligence or Amazon Bedrock
Data Automation, put retrieval on Foundry
IQ or Bedrock Knowledge Bases, and trace
everything in Langfuse.
## What I left out
This is client work under NDA. Not included,
on purpose:
- clients' names, their documents and their
data;
- deal details: proposals, pricing and
outcomes;
- prompts, model weights and fine-tuning data;
- POC names and the size of the team.
## How to verify
- **LinkedIn projects under Royal Cyber** (Dec
2022 – May 2024): "Enterprise LLM
fine-tuning, document search and chatbots"
and "Multimodal document intelligence"
(~40%).
- **A public LinkedIn post from June 2024**
describes leading a team of junior
engineers on GenAI proofs of concept and
leading pre-sales client conversations.
- **LinkedIn honors:** Star Performer of the
Year 2023, Royal Cyber.
- **Employment check:** Royal Cyber Inc.,
Senior Data Scientist, Dec 2022 – May
2024.
- **In an interview:** a walk-through of the
method, within NDA limits.
- **Code:** walkthroughs available on request.
Case study 10 · Royal Cyber Inc. · Dec 2022 – May 2024
Enterprise GenAI, from sales call to working POC
Fine-tuned LLMs, document search, multimodal document AI, and the client conversations that scoped them
Turning "can AI do this?" into scoped, shipped work: fine-tuned LLMs, enterprise document search, and multimodal document AI that cut manual review time by ~40%.
✓ Compiled from this page's own markdown.
Context
In 2023 every client asked the same question: "Can AI do this?" At Royal Cyber I was often the one answering it. First on pre-sales calls with clients in North America, Canada and the Middle East, then with a proof of concept, then with the build.
I was a senior data scientist there from December 2022 to May 2024. Before I moved on, I also led a team of junior engineers building GenAI proofs of concept.
The problem
Most GenAI requests arrive as a wish, not a spec. "Read our documents." "Answer our customers." "Summarize the contracts." The job is to find the part that's real, prove it fast, and be honest about the rest.
Constraints
- Enterprise documents in legal, financial and customer-support settings: messy scans, tables and images, not clean text.
- Cloud and on-premises deployments, depending on the client.
- The 2023 model landscape: GPT, LLaMA, Falcon and Mistral, with limited GPU budgets. That made parameter-efficient fine-tuning the practical route.
- Proofs of concept had to persuade. A demo that only works on the happy path doesn't survive a client's first real document.
What I built and did
Pre-sales. I took part in, and then led, pre-sales client calls: listening, finding the actual problem, proposing a tailored solution, negotiating, and keeping the technical plan tied to the client's goals.
Fine-tuned LLMs. I fine-tuned GPT, LLaMA, Falcon and Mistral models with LoRA and PEFT, then put them to work:
- enterprise document search using dense-retrieval RAG on FAISS and Pinecone;
- summarization for legal and financial documents;
- customer-support chatbots.
Multimodal document intelligence. Systems that read documents using text, vision and structured data together, so people review only what the model is unsure about. Manual review time fell by ~40%. The same extracted content fed multimodal RAG across text, images and tables.
A small team, real POCs. I led a team of junior engineers building GenAI proofs of concept for real-time applications, and explained their value to clients in plain language.
How it fits together
Scope first. A pre-sales call turns a wish into a scoped proposal and a proof-of-concept plan, and the team builds the POC.
For document use cases, documents (scans, PDFs, tables, images) go through extraction: OCR, layout and tables. Text and vision models classify each document and pull out its fields. Anything below a confidence bar goes to a human; everything else flows on. Extracted content is indexed for multimodal retrieval, and fine-tuned models handle domain language for summaries and support. The result shows up as document search, summaries and chatbots that answer with sources.
Stack
- Models: GPT, LLaMA, Falcon, Mistral; fine-tuning with LoRA and PEFT; Hugging Face Transformers; PyTorch, TensorFlow and Keras.
- Retrieval: dense-retrieval RAG and multimodal RAG; FAISS, Pinecone; embeddings.
- Documents: OCR, layout and table extraction, vision models.
- Platform: Python, Microsoft Azure (Azure Machine Learning, Azure Functions, Azure Kubernetes Service), cloud and on-premises deployment.
Results
- ~40% less manual review time with multimodal document intelligence.
- Pre-sales calls led across North America, Canada and the Middle East.
- A team of junior engineers led through GenAI proofs of concept.
- Star Performer of the Year 2023 at Royal Cyber.
- No recorded figures for the fine-tuning, search or chatbot work, so there are none here.
What I'd do differently
- Write the eval set in the first call. The fastest way to scope a GenAI project is to ask for twenty real examples and what "right" looks like for each. Today that's the first deliverable, before any model.
- Retrieve better before fine-tuning. In 2023, fine-tuning felt like the serious option. Most of those wins now come from better retrieval, structured outputs and a tight eval loop. Fine-tuning comes after, if at all.
- Measure reviewer time from day one. The ~40% came from reviewing less. Timing reviewers per document from the first pilot makes that kind of number airtight.
- Use the 2026 managed pieces. For extraction, I'd now start with Azure AI Document Intelligence or Amazon Bedrock Data Automation, put retrieval on Foundry IQ or Bedrock Knowledge Bases, and trace everything in Langfuse.
What I left out
This is client work under NDA. Not included, on purpose:
- clients' names, their documents and their data;
- deal details: proposals, pricing and outcomes;
- prompts, model weights and fine-tuning data;
- POC names and the size of the team.
How to verify
- LinkedIn projects under Royal Cyber (Dec 2022 – May 2024): "Enterprise LLM fine-tuning, document search and chatbots" and "Multimodal document intelligence" (~40%).
- A public LinkedIn post from June 2024 describes leading a team of junior engineers on GenAI proofs of concept and leading pre-sales client conversations.
- LinkedIn honors: Star Performer of the Year 2023, Royal Cyber.
- Employment check: Royal Cyber Inc., Senior Data Scientist, Dec 2022 – May 2024.
- In an interview: a walk-through of the method, within NDA limits.
- Code: walkthroughs available on request.