← Back to About
Career Arc

From writing business rules
to orchestrating autonomous agents

Fifteen years. Five phases. Each one built the intuition the next one needed — domain logic, statistical modeling, systems architecture, graph reasoning, and now LLM orchestration under regulatory constraint.

01
2009 — 2014 · Cognizant · Chennai → Gurgaon
The enterprise apprentice
Programmer Analyst → Product Specialist · 5 years
Wrote code against business specs for half a decade. Not glamorous — but you can't architect systems that survive regulated industries unless you've lived inside their logic first. This phase installed the domain intuition that every later phase would draw on. Went from implementing specs to defining them.
Enterprise logic SQL Business analysis Product ownership
The shift

Stopped asking "what does the spec say" and started asking "what does the data say." Pivoted out of software engineering entirely.

02
2014 — 2016 · EXL · Noida
The model builder
Assistant Manager — Data Science · 2 years
First pure ML role. Built, validated, and retrained predictive models across the full lifecycle — regression, classification, feature engineering from raw enterprise data. Learned the hard lesson early: a model without a retraining pipeline is a ticking clock. Accuracy decays. Distributions drift. The pipeline is the product.
Predictive models Feature engineering Model lifecycle R / Python
03
2016 — 2019 · Cognizant → Nagarro · Gurgaon
The signal hunter
Associate → Associate Lead · 3 years
Applied ML to real revenue problems — mortgage refinance targeting, order fallout root-cause analysis, churn reduction. Then NLP: mined unstructured product data at scale to extract sentiment, causal drivers, and latent themes — turning raw customer voice into structured signals that informed marketing and product strategy. Built propensity-to-expand models for IoT device company. First time owning the full arc from raw data to business recommendation.
NLP / Topic modeling Sentiment analysis Mortgage / FinTech Customer segmentation Random Forest
The shift

Moved to the US. The question changed from "can we predict X" to "how do we deploy, govern, and audit this at enterprise scale?"

04
2019 — 2023 · Incedo · Pittsburgh
The systems architect
Senior → Lead Data Scientist · 4 years
The model-builder became a systems-thinker. Built real-time fraud detection — dual-layer architecture with a classifier feeding an autoencoder. Deployed credit scoring via Kafka messaging for a major bank's card portfolio. Designed COVID-era risk adjustment when traditional models broke overnight. Then the graph era — designed ontologies, built RDF knowledge bases, ran vendor evaluations for graph databases, and championed data fabric architecture.
Kafka / Real-time Fraud detection Knowledge graphs Ontology design Credit scoring AWS Data Fabric
The shift

LLMs arrived. The knowledge graph work and NLP instincts converged — how do you make these models work under regulatory constraints, with audit trails, at production scale?

05
2023 — present · Incedo + Independent · Pittsburgh
Current phase
The orchestrator-researcher
Senior Lead Data Scientist + Independent Researcher
Dual track. By day: architecting multi-agent LLM orchestration on AWS Bedrock with Neo4j knowledge graphs for regulatory compliance — RAG pipelines, automated retraining, data quality governance across the full ML lifecycle. By night: formalizing in-context learning theory through an information-theoretic lens, studying memorization distribution across transformer layers on A100, and deconstructing GPTQ from scratch on 7B models. Publishing the internals — what the papers don't tell you, what actually breaks at scale.
Multi-agent / MCP AWS Bedrock RAG Neo4j MLOps ICL Theory GPTQ PyTorch / A100

Each phase was prerequisite to the next. You can't orchestrate compliance AI agents without having first lived inside enterprise logic, then built the models, then architected the systems, then understood the graphs. The stack compounds.