Five engagements, written up the way I’d want to read them.

Each one follows the same spine: context, problem, constraints, approach, outcome, and what I’d do differently. The constraints section is usually the interesting one.

cdc: 50 source systems → customer.materializedordering: per customer
Rocket India, US mortgage servicing, 2021-2025

Unifying 500+ tables from 50+ source systems into one real-time customer view

A streaming customer data platform built without a freeze window on any source system, serving 1M+ API calls at peak hour in real time, with a Snowflake gold layer that cut a common downstream query from about three seconds to roughly 500 milliseconds.

500+ tables unified50+ source systems1M+ API calls at peak hour3s to 500ms query latency
KafkaConfluentDebezium CDCksqlDBSnowflakeKubernetesJava
Read the full case study
Rocket India, Kafka platform, 2019-2026

Making Kafka self-service for 200+ teams without letting any of them near production data

One control plane where naming, retention, ownership and access are enforced at request time rather than agreed in a document, still in production and governing 5,000+ topics.

5,000+ topics governed200+ teamsone writer per topic, alwaysproduction closed to self-service
KafkaConfluentKubernetesFlinkSchema RegistryKafka ConnectJava.NET
Read the full case study
Rocket India, Confluent Cloud migration, 2025

Migrating a production Kafka estate to Confluent Private Service Connect

Every producer and consumer moved to private networking incrementally, team by team, with zero data loss and no re-keying of existing topics, alongside a new real-time data exchange platform for secure, cross-cluster sharing.

zero data lossincremental, team-by-team cutover200+ teams on the new standardno topic re-keying
Confluent CloudPSCCluster LinkingTerraformKafkaGCP
Read the full case study
Mintly, hiring platform, 2022-present

Building a hiring platform where agentic work stays offline and live recommendations stay deterministic

A nine-service platform serving 1,000+ employers and 20,000+ jobseekers, where agentic matching work runs entirely offline and the live recommendation path never calls a model.

9 production services1,000+ employers20,000+ jobseekerszero LLM calls on the live path
Spring WebFluxGraphQLNext.jsFastAPILangGraphMongoDB Atlas Vector SearchAWSTerraform
Read the full case study
ResumeVera, a product I own end to end

Shipping a public product alone, from the model call to the deployment pipeline

A live product doing three jobs, a resume builder, an ATS check and a LinkedIn optimiser, with its own API, front end and deployment pipeline, built and run by one person.

solo buildresume, ATS check, LinkedInAPI, front end and infrastructurein production
Next.jsNodePostgreSQLCloud Run
Read the full case study