The work
Production systems running in global financial workflows. Plus personal projects exploring the frontier of GenAI and LLMOps.
Production work at LSEG
Financial AI systems delivering measurable outcomes in global workflows
GenAI Financial Data Extraction Pipeline
£350K+ saved · 50 FTE reduction
Meta-prompting and PydanticAI for extracting structured financial data from broker reports. The accuracy came from deeply understanding the document format — not just the model choice.
Context-Aware RAG Answer Engine
Source-referenced · Session memory · Analyst-grade
RAG system built specifically for financial analysts. Source-referenced because compliance requires auditability. Session memory because analyst workflows are multi-turn. Prompt design informed by knowing exactly how analysts phrase financial queries.
Korean Regulatory Filing Extraction
15 days → 40 minutes · 130K+ documents · £250K+ saved
End-to-end extraction for Korean financial regulatory filings. 85%+ hit rate across 130K+ documents, 5 FTE reduced. The speed improvement wasn't the hard part — the precision in a non-English financial context was.
Multi-Agent Financial Document Intelligence PoC
Executive-endorsed · LangGraph orchestration
Multi-agent PoC for intelligent financial document processing. Designed the agent architecture knowing the document types, the financial context, and the analyst workflows it needed to serve. Earned regional executive endorsement at LSEG.
Image-Based Document Intelligence Pipeline
0% → 64% image coverage · 5 FTE eliminated
Two-stage pipeline: Azure Layout Model for structure extraction + Python rule-based processing for financial semantics. Expanded coverage of image-based financial documents from zero to 64%, opening an entirely new automation surface.
Hybrid ML Ingestion Pipeline
5 markets · 3 FTE saved · 50% scope reduction
Hybrid cosine similarity + two-stage classification pipeline for financial document ingestion across CA, IN, US, AU markets. Rule-based edge cases handled by someone who'd actually seen those edge cases in production financial data.
Personal projects
Exploring the frontier on personal time. All on GitHub.
Smart Note Chrome Extension
LLM-powered YouTube note-taking with real-time video summarisation. Built to solve a personal pain point — and shipped as a Chrome extension.
LLaMA 3.1 8B Fine-tuning
Domain-adapted LLaMA 3.1 8B via Unsloth QLoRA — ~60% faster inference than standard PEFT while maintaining task quality. The dataset quality mattered more than the training config.
MLOps Query Classifier
End-to-end MLOps pipeline with CI/CD, CML, drift monitoring, and automated retraining. A demonstration of what production-grade ML lifecycle management actually looks like.
BERT from Scratch
MLM pre-training of BERT from scratch, fine-tuned on Stanford Sentiment Treebank. End-to-end understanding of the transformer training loop.
GenAI Leadership Recognition
LSEG · 2023
Multi-agent AI PoC earned regional executive endorsement.
Hall of Fame Award
LSEG · 2023
Recognised for Non-English market automation delivering measurable impact.
Speaker — Learning Forum
LSEG · 2023
Presented on Python problem-solving to the internal engineering community.
PG Programme in Data Science
INSOFE / Carnegie Mellon University
Machine Learning in Production
Andrew Ng · Coursera
Complete MCP Developer Guide
Udemy
Six Sigma Green Belt
KPMG
Want to build something together?
abhilash.k.panda@gmail.com