Dipankar Banik

Senior Backend and Platform Engineer at Narvar

I build and run backend platforms: Java, Python, and Go services, event pipelines on Apache Pulsar and Kafka, and the Kubernetes and Terraform setup they run on. For the last ten years that has meant the backend of Narvar's Returns platform, which hundreds of retail brands use to handle their returns. Lately that includes LLM-powered agent workflows for fraud prevention and the Customer Data Platform, with a policy engine making the final, auditable call.

Experience
20+ years
Core stack
Java 21, Python, Go, Spring Boot, FastAPI
Cloud
GCP, AWS, Kubernetes
Work status
US permanent resident, no sponsorship needed

About

I've been writing software professionally since 2001. The first half of that was enterprise Java: portals, workflow systems, and integration work for Cisco, Kaiser Permanente, Gilead Sciences, Major League Baseball, and Marsh. A lot of it was getting new systems to work with old ones, such as AS400 and DB2 mainframes, Oracle Access Manager single sign-on, and Salesforce.

Since 2016 I've been at Narvar, working on the backend of the Returns platform. I own services end to end: design, code, infrastructure, and on-call. The problems I enjoy most are the messy middle of distributed systems, like carrier webhooks that arrive late, twice, or out of order, where the system still has to end up with the right answer.

Recently I've been working on returns fraud prevention and the Customer Data Platform. In both, a Drools policy engine makes the final decision so it stays auditable and repeatable, and multi-step LLM workflows (MCP tool calling, RAG, vector search) gather and summarize the signals that are hard to write as rules. On my own time I build agent systems in Python end to end: multi-agent orchestration on LangGraph and LangChain with human-in-the-loop steps, FastAPI services routing across OpenAI and Anthropic models, RAG tuned against evaluation datasets, LLM-as-judge regression tests in CI, OpenTelemetry tracing, and guardrails for prompt injection and PII.

I've led and mentored teams of six to ten engineers across the US and India. At Narvar I set the engineering standards for quality, security, and observability, lead architecture, design, and pull-request reviews, and partner with product and security teams. Away from the keyboard I build small Raspberry Pi projects (gpati is on my GitHub) and walk the trails on Pleasanton Ridge.

Experience

Mar 2016 – Present

San Francisco, CA

Narvar

Senior Backend and Platform Engineer, Returns Platform

Narvar's Returns platform handles returns, exchanges, and refunds for hundreds of retail brands.

  • Own the backend of the returns SaaS: Java 21 and Spring Boot 3 microservices on GCP Kubernetes, backed by Postgres, AlloyDB, Redis, and Apache Pulsar. Led the production migration from AWS to GCP.
  • Built the carrier event pipeline on Pulsar for UPS, FedEx, USPS, and Newgistics, with shared subscriptions and dead-letter queues, plus a reconciliation layer that keeps each shipment's event history correct when webhooks arrive late, duplicated, or out of order.
  • Wrote the return eligibility rules engine: Drools with a custom DSL parsed by ANTLR, deployed as 38 Java Lambdas, so retailer-specific rules can change without a code release.
  • Built a webhook return-tracking pipeline in Go with Cloud Tasks scheduling and dead-letter handling.
  • Maintain the infrastructure as code: Terraform across AWS and GCP, 79 Helm charts, ArgoCD, Ansible, and Packer (300+ commits). Services scale from 3 to 99 pods at peak.
  • Handle service security with OAuth2 and Okta, mTLS through Linkerd, and secrets in HashiCorp Vault. Own the Google Tink based PII encryption library used across the platform, which was also part of our log4j2 remediation.
  • Fraud prevention: built the returns fraud prevention platform. A Drools policy engine with a custom ANTLR DSL makes the final, auditable decision; multi-step LLM workflows (MCP tool calling, RAG, vector search) gather and summarize the signals, so results are repeatable and explainable.
  • Customer Data Platform: own the backend and data governance for the CDP: Python tooling for pipeline and schema governance (84 Avro and 23 Protobuf schemas) feeding a 494-table BigQuery warehouse, and rule-guarded AI workflows with policy checks on every LLM output.
  • Shipped Returns V2, Sales Assist, Boxless Returns, Simple Returns, and the Stripe payments integration.
  • Technical leadership: set the engineering standards for quality, security, and observability; lead architecture, design, and pull-request reviews; mentor engineers; partner with product and security teams; write the design docs and ADRs; take part in the on-call rotation.

Java 21 · Spring Boot 3 · Python · Go · Apache Pulsar · Kafka · Postgres · AlloyDB · Redis · BigQuery · Drools · LLM agents · MCP · RAG · Kubernetes · Terraform · Helm · ArgoCD · GCP · AWS

Sep 2014 – Mar 2016

Foster City, CA

Gilead Sciences

Applications Engineer

  • Led the migration of the Product Quality System from .NET to Java EE on Liferay 6.2. The system tracks quality records through the drug lifecycle, is FDA-validated, and runs at four sites: Foster City, San Dimas, Oceanside, and Edmonton.
  • Built the JSR-286 portlets, a Spring MVC and Hibernate service layer, LDAP authentication hooks, and a data access layer that works against both Oracle and SQL Server.

Java EE · Liferay 6.2 · Spring MVC · Hibernate · Oracle · SQL Server · LDAP

Aug 2010 – Aug 2014

San Jose, CA

Cisco

Portal Lead

  • Led GTMS uCRM, the Cisco.com customer support portal on WebSphere Portal 6.1, with Salesforce as the case management backend. Designed the Salesforce integration and the WSRP portlet catalog.
  • Built CVCM portlets on Cisco's Liferay-based QUAD platform, giving collaboration users access to more than 200 tools with role-based search and real-time event subscriptions.

WebSphere Portal 6.1 · Liferay · Salesforce · Spring MVC · JSR-286 · WSRP

Oct 2009 – Aug 2010

Pleasanton, CA

Kaiser Permanente

Lead WebSphere Portal Analyst

  • Designed and built UBT Tracker, the HR platform for Kaiser's Unit Based Teams: team setup, roles, goals, and labor contract tracking. WebSphere Portal 6, Spring 2.5, Hibernate JPA, single sign-on through Oracle Access Manager, and Spring Batch ETL from DB2 mainframes.

WebSphere Portal 6 · Spring · Hibernate JPA · Spring Batch · DB2 · Oracle Access Manager

2005 – 2009

New York, NY and Hoboken, NJ

Consulting engagements

Lead Portal Architect and Developer

  • Major League Baseball: designed the Financial Desktop portal framework on Struts 2 and WebSphere Portal 6.1 and led a team of six using test-driven development.
  • Marsh USA: technical lead for MarshConnect, Marsh's global client portal, with a team of ten.
  • Cravath, Swaine & Moore: migrated the firm's portal from WebSphere Portal 5.1 to 6.1, including content and WSRP portlets.
  • Country-Wide Insurance: built a broker portal that pulled together AS400 and DB2 data.
  • Guy Carpenter: built a real-time straight-through-processing reporting platform and extended Apache POI to embed images in JasperReports Excel exports.

2001 – 2005

Various

Early career

Programmer and J2EE Developer

  • Java and J2EE systems on EJB 2.0, WebLogic, and Oracle for John Deere (including its Asset Information System) and Marsh. Before that, handheld sync for Pocket PC and Palm in J2ME, a logistics content management system, and a remote security monitoring platform.

Selected work

A few problems I've spent real time on, and how I approached them.

Carrier events that arrive in the wrong order

Carriers send tracking webhooks late, twice, or out of sequence. I built a Pulsar pipeline with per-shipment ordering and a reconciliation step that rebuilds each shipment's timeline from what actually arrived, so customers see the right status even when the input is messy.

Apache Pulsar · Java 21 · Postgres

Return rules that change without a deploy

Every retailer has its own return policy. I built an eligibility engine on Drools with a custom DSL parsed by ANTLR and deployed it as 38 small Lambdas, so business teams update rules without waiting for an engineering release.

Drools · ANTLR · AWS Lambda

Fraud prevention you can explain

For returns fraud, an LLM answer on its own isn't enough because every decision has to be explainable. A Drools policy engine makes the final decision, and multi-step LLM workflows (MCP tool calling, RAG, vector search) gather and summarize the signals, so results are repeatable and auditable.

Drools · LLM agents · MCP · RAG · vector search

A customer data platform with guardrails

Narvar's Customer Data Platform brings retailer and shopper data into a 494-table BigQuery warehouse. I own the Python tooling for pipeline and schema governance (Avro, Protobuf) and designed its rule-guarded AI workflows, where policy checks sit around every LLM output before anything acts on it.

Python · Avro · Protobuf · BigQuery · LLM workflows

One way to encrypt personal data

I own a Google Tink based client library for encrypting PII, with Resilience4j retries. Services across the platform use it instead of rolling their own, and it was part of how we remediated the log4j2 vulnerability.

Google Tink · Resilience4j · Java

Agent systems, built end to end

Independent work in Python to go deep on the agent stack: multi-agent orchestration on LangGraph and LangChain with human-in-the-loop approval steps and tools exposed through MCP, FastAPI services routing across OpenAI and Anthropic models with caching and cost tracking, RAG tuned against evaluation datasets, LLM-as-judge regression tests in CI, OpenTelemetry tracing, and guardrails for prompt injection and PII.

Python · LangGraph · FastAPI · MCP · OpenTelemetry

Skills

Languages
Java 21, Python 3.12 (FastAPI, asyncio, Pydantic), Go, SQL, TypeScript, Kotlin
Backend
Spring Boot 3, Hibernate, Spring Batch, Drools, ANTLR, Quartz, Resilience4j
Messaging and APIs
Apache Pulsar, Kafka, Cloud Tasks, REST, gRPC, GraphQL, webhooks
Data
Postgres, AlloyDB, Yugabyte, Redis, BigQuery, Oracle, DB2, Avro, Protobuf
Cloud and infrastructure
GCP, AWS (Lambda, S3, VPC, ECS), Kubernetes, Docker, Terraform, Helm, ArgoCD, Linkerd, Ansible, Packer
Security
OAuth2, Okta, JWT, mTLS, HashiCorp Vault, Google Tink, ModSecurity
Testing
JUnit 5, Mockito, Cucumber, Playwright, Jest
Agents and LLMs
LangGraph, LangChain, MCP, multi-agent orchestration, human-in-the-loop, OpenAI and Anthropic APIs, RAG, vector databases, LLM-as-judge evals, guardrails, OpenTelemetry tracing
Frontend
React, TypeScript, Next.js, Apollo GraphQL
Leadership
Teams of six to ten across the US and India, mentoring, engineering standards, architecture and design reviews, incident reviews

Companies I've built software for

  • Narvar
  • Gilead Sciences
  • Cisco
  • Kaiser Permanente
  • Major League Baseball
  • Marsh
  • Guy Carpenter
  • John Deere
  • Cravath, Swaine & Moore
  • Country-Wide Insurance

Education

B.S. in Physics, University of Pune, 1999

Contact

Email is the quickest way to reach me. I'm interested in senior, principal, and technical lead roles in backend, platform, and AI agent systems, in the Bay Area or remote.