Hi, I'm Sripathi M.
Data & AI Architect. I build the layer that tells enterprise AI what the data means.
Ten years in engineering, seven in data, the last four at Fractal - enterprise data architecture, governance and data modelling on Azure and Databricks, across supply chain, consumer goods and retail. Today I architect the logistics ontology and the semantic layer beneath it for a Fortune 500 CPG company.
Ask three teams in a large company for their on-time delivery figure and you get three different answers - each correct by its own definition. People resolve that ambiguity in a meeting. An AI agent cannot: given inconsistent definitions it returns a confident answer that is simply wrong.
Most of my decade has gone into the layers underneath that problem - ETL and dimensional modelling, then lakehouse platforms, then the architecture and governance around them. I led a five-engineer team on a Databricks transportation platform of 300+ production tables. Its metadata-driven Medallion framework meant new source systems arrived as configuration, not as new pipeline code.
I also build AI agents for data engineering workflows - including a voice agent that runs first-round SQL interviews.
My career started in electrical and power-plant engineering. That background still shapes how I think about data: a wrong reading has a real cost.
Platform I led
300+ production tables on a metadata-driven Medallion framework
Team
Five engineers across the transportation data platform
Governance
Every KPI definition reconciled against source data before release
Progression
Two promotions in four years at Fractal
CURRENT WORK
Architect - Data & AI, Ontology & Semantic Layer
For the logistics domain of a Fortune 500 CPG company, I model shipments, loads, carriers and routes as object types with explicit relationships between them - then define business KPIs as governed measures over that model, rather than as concepts sitting alongside it. The semantic layer maps the model onto the physical transportation tables, so one business question resolves to one governed definition.
No definition ships until it has been recomputed from source and tied out against the figure the business already trusts. That reconciliation gate is the part I would defend hardest - it is the difference between a model agreed in a workshop and one that matches the data underneath it.
- Ontology
- The object and relationship model for the logistics domain
- Semantic layer
- The mapping from that model onto the physical transportation tables
- Governance
- Every governed KPI reconciled against source data before release, with a named owner per definition
- Consumers
- LLM agents and analytics resolve questions against governed concepts instead of raw tables
- Stack
- Timbr (SQL ontology over Databricks) · Databricks · Azure · PySpark · SQL
PORTFOLIO
BLOG
Semantic Search, Lesson 1: Why Keyword Search Only Gets You Halfway
Notes from DeepLearning.AI's Large Language Models with Semantic Search course - how BM25 keyword search actually works, why it breaks on paraphrased queries, and where embeddings, dense retrieval, and rerank fit into the bigger RAG picture.
NLP Basics: Bag of Words, Text Preprocessing, and a Real Kaggle Walkthrough
Notes on how NLP turns text into numbers a model can use - Bag of Words, tokenization, stemming vs. lemmatization - worked through end-to-end on the Quora Insincere Questions Classification dataset.
Docker Notes: From Virtual Machines to Containers
My learning notes on why Docker beat traditional VMs, how its architecture evolved from a single daemon to containerd + shim + runc, and the basics of images, registries, and the CLI.