I make sure the data behind AI systems actually holds up in the real world.
I've spent the last three years elbow-deep in data pipelines — the unglamorous plumbing most people never think about until it breaks. Somewhere along the way that plumbing started carrying AI and ML workloads, and that's the part of the job that stuck with me.
These days I spend my time making sure the data feeding machine learning models is fast, clean, and something you can actually trust — whether that's a real-time safety alert or a chatbot answering questions in plain English.
Real-time detection that replaced a rulebook
Built the pipeline that feeds live sensor data from 20+ industrial wells into a machine learning model, replacing a rigid rules-based system with one that actually learns.
5–6 min alert latencyA data foundation moving 100GB a day
Designed ETL pipelines pulling from APIs, databases, files, and event streams into a medallion architecture, feeding downstream teams doing analytics and AI.
100GB/day · 6+ consumersCutting cost without cutting corners
Rethought how messages were batched and deduplicated in an event-driven system, trimming both cloud spend and end-to-end latency.
−60% cost · −30% latencyProtecting proprietary logic at scale
Migrated production pipeline modules to a Cython-based architecture, securing IP across multi-tenant deployments without touching Spark performance.
40+ modules securedTurning deployment into a non-event
Set up Databricks Asset Bundles as the standard deployment framework, embedding Git-based release management across the org.
−80% deploy timeKeeping history accurate under audit
Implemented SCD Type 3 and CDC patterns for large-scale historical backfills, keeping data accurate across audited customer environments.
SOC2 auditedTeaching a chatbot where its limits are
Built a conversational analytics assistant with a multi-agent routing layer and prompt templates per data domain, so it only ever answers from data it's allowed to see.
guardrailed · multi-agentStandardizing how questions get asked
Designed prompt templates per domain to keep question framing consistent, improving routing accuracy across the whole conversational system.
per-domain prompt designA B.Tech in Computer Science and Business Systems from VIT — the overlap in that degree title turned out to be more prophetic than I expected, since most of my job now sits exactly at that intersection.
Currently working through the Databricks Certified Data Engineer Professional and AWS Certified Data Engineer – Associate exams, mostly because production has a way of teaching you what you assumed instead of what you verified.