Abhilash
Kumar Panda
I spent two years inside institutional finance reconciling pension funds.
Most AI engineers call that irrelevant.
I call it an unfair advantage.
"AI systems in finance fail when built by engineers who've never lived inside financial markets. I've done both — and that's the difference between AI that demos well and AI that delivers £600K+ in production."
From the desk to the model layer
Three chapters. Each one built on the last.
Reconciliation Analyst
State Street Global Advisory
100+ pension fund portfolios. Bloomberg Terminal, Refinitiv Eikon, SmartStream TLM. Daily reconciliation across equities, derivatives, and FX. This became the domain knowledge that made every AI system at LSEG actually work.
- 100+ managed pension fund portfolios — equities, derivatives, FX, corporate actions (M&A, dividends)
- Built PowerBI exception-trend dashboard — frontline data pain that directly shaped LSEG AI automation
- Cross-custodian workflows, post-trade settlement, and regulatory reporting expertise
Data Scientist
London Stock Exchange Group
Turned domain expertise into production ML pipelines. 130K+ documents, 15 days → 40 minutes, £250K+ saved. Proof that knowing the domain first makes better ML engineers.
- Korean regulatory filing extraction: 85%+ hit rate, £250K+ savings, 5 FTE reduction, 130K+ documents
- Hybrid ML + rule-based ingestion across CA, IN, US, AU — 3 FTE saved, 50% out-of-scope reduction
- Azure Layout Model pipeline — image-based document coverage 0% → 64%, eliminating 5 FTEs
Senior Data Scientist
London Stock Exchange Group
GenAI extraction pipelines, context-aware RAG, multi-agent PoC earning executive endorsement. £350K+ delivered. Commercially accurate because financial domain knowledge is embedded from day one.
- GenAI financial data extraction pipeline (meta-prompting, PydanticAI) — £350K+ savings, 50 FTE reduction
- Context-aware RAG Answer Engine with source-referenced responses and session memory for analysts
- Multi-agent AI PoC for financial document intelligence — earned regional executive endorsement
What this looks like in production
Not demos. Systems running in global financial workflows.
Financial Data Extraction Pipeline
£350K+ saved · 50 FTEs automated
Meta-prompting + PydanticAI for structured extraction from broker reports. Accurate because the prompts were designed by someone who has read hundreds of these reports.
Context-Aware Answer Engine
Source-referenced · Session memory
RAG for financial analysts — source-referenced for compliance auditability, session memory for multi-turn workflows, prompt design informed by actual analyst query patterns.
LLaMA 3.1 8B Fine-tuning
~60% faster than standard PEFT
Domain-adapted via Unsloth QLoRA. Dataset quality was fundamentally different because the financial documents were understood, not just collected.
What I work with
9 years of tools — selected for production, not benchmarks.
GenAI & Agents
LLMs
ML / Deep Learning
Cloud & LLMOps
Seeking my next chapter
Nine years in. Looking to bring financial AI expertise to Singapore's fintech ecosystem.
Senior IC roles, lead positions, advisory engagements where deep financial AI expertise matters.