AI/ML Engineer · Researcher · Pittsburgh

Building systems
that reason
at scale

I design and ship production LLM infrastructure, graph analytics systems, and compliance-grade AI pipelines — then publish the internals.

5+ Production Systems
3 Research Papers
AWS Bedrock · EKS · EC2
Neo4j Graph Analytics
01

About me

I'm a Data Scientist / Data Engineer working at the intersection of large language models, graph analytics, and enterprise compliance. My day job involves building production AI systems on AWS infrastructure — the kind that need to be auditable, explainable, and robust under regulatory scrutiny.

My current focus is Project Kratos, an LLM-based multi-agent system built on AWS Bedrock, EKS, and Neo4j. Outside of that, I'm independently researching in-context learning theory and LLM quantization — writing papers I intend to publish on arXiv.

I write about the gap between ML research and production reality: what the papers don't tell you, what actually breaks at scale, and what's worth building from scratch.

Engineering
LLM Agents AWS Bedrock Neo4j Kafka Spark Python
Research
In-Context Learning GPTQ Memorization Transformers
Domain
Banking / FinTech Fraud Detection Regulatory AI GRC
02

Selected Projects

LLM Infrastructure

Project Kratos

Multi-agent LLM system on AWS. Orchestrates specialized agents across Neo4j knowledge graphs and Bedrock model inference for enterprise compliance workflows.

AWS Bedrock EKS Neo4j MCP Python
Data Quality

GX Forge

Streamlit tool converting Oracle SQL data quality rules into Great Expectations YAML. Handles compound queries, SSL proxy environments, and an 8-module rationalized architecture.

Python Streamlit Great Expectations Oracle SQL
Graph Analytics

Fraud Graph Engine

Neo4j-based fraud detection using community detection algorithms (Louvain, k-core, label propagation) and structural entropy for transaction network analysis.

Neo4j GDS Louvain Cypher Python
Compliance AI

FDIC Part 370 Framework

110-control IT compliance framework with agentic 8-pass LLM extraction pipeline. Covers FDIC Part 370 and 12 CFR Part 330 with full regulatory basis mapping.

Claude API Cytoscape.js GRC FDIC
Stream Processing

Security Event Detector

Spark streaming application for Kafka-based compound security event detection (password + address + card changes in same session) with YAML-driven pattern configuration.

Spark Streaming Kafka YAML Scala
NLP Pipeline

Relationship Extractor

Multi-pass relationship extraction system using Claude API with pronoun resolution and iterative refinement across unstructured text corpora.

Claude API NLP Python Neo4j
03

Research Work

2025

An Information-Theoretic Framework for In-Context Learning

Formalizes ICL success conditions: mutual information between prompt features and output correctness must exceed task entropy. Derives tractable bounds and empirical predictions.

In Progress
2025

Layer-wise Memorization Sensitivity in Language Models

Empirical study of memorization patterns across transformer layers using GPT-2 on A100 GPU. Identifies which layers disproportionately contribute to verbatim memorization.

In Progress
2024

GPTQ Quantization from Scratch on 7B Models

Reimplementation and analysis of GPTQ on 7B-parameter models. Documents per-layer quantization error accumulation and accuracy/compression tradeoffs.

Draft
04

Technical Writing

Why Your LLM Agent Needs a Gateway Before Anything Else

Context management strategy for multi-agent systems. The LLM gateway as foundational infrastructure — cost control, observability, rate limits, and model switching.

Structural Entropy as a Fraud Signal in Transaction Networks

How graph entropy metrics catch money mule networks and synthetic identity fraud that rule-based systems miss entirely.

The Lottery Ticket Hypothesis: What It Actually Implies for Production Models

Beyond the paper — practical implications of sparse subnetwork theory for fine-tuning, pruning, and deployment at scale.

Building Compliance-Grade AI in Banking: What the Frameworks Miss

A practitioner's view of FDIC Part 370, OCC standards, and the gaps between regulatory intent and actual AI system behavior.

05

Tech Stack

LLM / AI
  • Claude (Anthropic)
  • AWS Bedrock
  • LangChain
  • MCP Protocol
  • Hugging Face
  • GPTQ / bitsandbytes
Data / Graph
  • Neo4j + GDS
  • Apache Spark
  • Kafka
  • Great Expectations
  • Oracle SQL
  • Cytoscape.js
Infrastructure
  • AWS EC2 / EKS
  • Docker / Kubernetes
  • Python / Scala
  • Streamlit
  • FastAPI
  • Git / GitHub
Research
  • PyTorch
  • A100 GPU (CUDA)
  • arXiv
  • Semantic Scholar API
  • GPT-2 / 7B Models
  • LaTeX

Let's connect

Open to research collaborations, technical discussions, and consulting on LLM systems, graph analytics, or compliance AI in financial services.

Email sumitasthana@outlook.com GitHub github.com/sumitasthana/pratyaxa LinkedIn linkedin.com/in/thatsumit