The hard part isn't the model — it's making the data underneath defensible. Four years on enterprise modernization, analytics platforms, and now Rust and TypeScript systems for AI workflows that hold up under audit.
LEGACY_MODERNIZATION
900+
Java modules → Spark + Beam
STREAMING_SCALE
20M/day
events into BigQuery
DATASETS_ONBOARDED
35+
financial datasets → lakehouse
INGEST_RUNTIME
6h→45m
119-table CDC + validation
01 / The work
Each one is a different angle on the same question: how do you make AI useful inside a real organization without breaking the things underneath?
pulseql.project
Governed analytics
product summary · public view
A governed data workspace for teams that want AI-assisted analysis without losing control of review, privacy, or operational trust.
atrium.project
Enterprise knowledge
product summary · public view
A knowledge product for teams that need answers from company documents with clear citations, honest refusal, and permission-aware user experience.
relay.project
Safer AI workflows
product summary · public view
A coordination layer for AI-assisted engineering teams that need shared context, stronger review signals, and safer workflows across tools.
Ground what it can. Refuse what it cannot. Run within bounds the team can defend.
02 / What I believe
That shorthand above sits on top of these three. Whether the system is a modernization, an analytics platform, or an AI workflow, the same primitives keep showing up.
Reliable AI starts with governed data systems: ingestion, schema enforcement, data quality, lakehouse tables, semantic definitions, and lineage.
Iceberg + BigQuery lakehouse work across 35+ financial datasets and 20+ years of reporting context.
LLM systems need retrieval evidence, metric contracts, access boundaries, audit trails, and explicit execution paths before they can be trusted.
RAG work over 100K+ survey responses reduced qualitative review effort by 60%.
Large rewrites need migration systems, not one-off scripts: repeatable transformations, validation gates, and clear ownership for every generated artifact.
GenAI-assisted modernization moved 900+ Java modules toward Spark and Apache Beam in roughly seven months.
Data engineering
Batch, streaming, quality, and lakehouse work for enterprise data teams.
Pipelines, semantic layers, and modernization systems at production scale.
ML & AI systems
Model workflows, feature systems, deployment paths, and evaluation loops.
Built across applied ML, MLOps, and AI-assisted review workflows.
Generative AI
Grounded LLM applications that keep context, evidence, and review visible.
RAG systems, knowledge interfaces, and agentic AI infrastructure.
Cloud & DevOps
Cloud delivery, infrastructure automation, observability, and platform hygiene.
AWS, GCP, containers, CI/CD, and production monitoring workflows.
03 / The stack
Four layers — data, ML, generative AI, cloud. The tools differ; the test is the same: does this workflow hold up under audit?
The work is data infrastructure before it is AI.
04 / The path
Four roles teaching the same lesson in different ways — that the infrastructure decisions outlast the systems they live inside.
012025
Data Engineer
Wells Fargo · via Capgemini America Inc.
Charlotte, NC
022023
Data Specialist
University of Maryland
College Park, MD
032021
Senior Software Engineer. Data Platform
Tiger Analytics
Chennai, India
042019
Intern & Software Engineer
Xenonstack Pvt. Limited
Chandigarh, India
Offline
Shot on an Olympus EM-10 between trips and walks. A quieter counterweight to the systems work: composition, patience, and noticing what the frame leaves out.
05 / Notes
Short technical notes published when there's something specific to say. The decisions behind the systems, in writing.
The shape of a local-first data product
Two storage decisions sit at the heart of any local-first analytics tool. Get them right and you ship as a single application the user can trust on their own machine. Get them wrong and you've reinvented a hosted service somebody has to operate.
Refusal is the feature grounded RAG actually needs
Trying to make the model better at admitting uncertainty is a losing game. Refusing in the orchestration layer. When the retrieval signal is too weak to support an answer. Is a winning one. Plus the companion move that matters at least as much.
Why AI should reason over metadata, not raw data
Schemas, contracts, lineage, and policies carry the structure an LLM needs to be useful inside a regulated environment. The bytes don't. And shouldn't leave the perimeter.
The case for local-first semantic layers
When the operator is the user, a single binary on their laptop beats a hosted control plane more often than the industry admits. Iceberg + DuckDB on hardware you already own does more than expected. And the trade-offs aren't where you'd guess.