Senior Backend Engineer & AI Practitioner

Systems that stay honest when things go wrong.

Backend engineer, 11+ years in. Most of that went into payment and data platforms: Kafka pipelines that could not drop or double-process a message, and a PostgreSQL estate that had to stay fast while it crossed 30 million rows. Lately the same habits go into AI systems. Retrieval, structured outputs, evals. The rest of this page is specifics.

Open to senior backend and applied AI roles · remote or relocation

Selected work

Five systems, 2013 to present.

No. 01

Fifteen million records, one canonical shape

HG Insights · 2021 to 2025 · geo-standardization service

The reports were quietly comparing two spellings of one city. I standardized 15M+ US and Canada company records into a single canonical country, state and city tree by building a geo-standardization service with shadow-writes and rollback, then rolling it out in phases with the EVP of Data Solutions. Later data-ingestion workflows were built on the same framework.

No. 02

The bad message

Capital One, via Cognizant · 2018 to 2021 · Kafka, financial transactions

A payments platform handles messages it cannot afford to lose or process twice. I hardened it against both failure modes by building a dead-letter-queue strategy with bounded retry, backoff, and idempotent replay. The test I checked against was simple: a re-run should not double-charge anyone. The metric I watched most was backlog per partition. One poison message at the head of a partition stalls everything behind it.

Fig. 1 · One poison message, set aside and replayed exactly once.

No. 03

Fast past thirty million rows

HG Insights · PostgreSQL · partitioning, indexing, replicas

The heaviest dashboard was slow and getting slower. I cut its latency by more than 60% on a 30M+ row PostgreSQL dataset by range-partitioning on region, rebuilding indexes around the queries people actually ran, and routing reporting reads to a replica. It stayed fast as the data grew. That was the actual requirement.

No. 04

Software a physical plant runs on

Agilent Technologies · 2026 to present · manufacturing

Software Engineer, Manufacturing (Backend & Automation). I build and fix the processes and recipes a production plant runs on. Separately, I build internal applications that put AI to practical use, and I lead the team's internal AI enablement. It is internal work, so I describe it by outcome. A wrong recipe here has a physical cost, which changes how carefully you ship.

No. 05

Holding AI work to the same standard

applied AI · LangGraph, RAG, structured outputs

Grounding and structured outputs are the queue semantics of AI systems: an answer should not be invented, and it should not lose the evidence behind it. Assay, a LangGraph multi-agent research tool I built, fans one manager out to four specialists in parallel, and every step emits structured JSON, so an answer traces back to what was retrieved. When evidence is thin it says so.

Side projects

Running in public.

Each link opens the real system.

Clarion live · early build

Complaint intelligence over the public CFPB database: vague free-text complaints become standardized, evidence-backed failure modes. Designed grounded or silent, so a classification cites its evidence or abstains.

Elixir · Phoenix LiveView · Broadway · Kafka · PostgreSQL + pgvector · Amazon Bedrock

signal.prateekmulye.dev ↗

ChatFormula1 live

An agentic RAG assistant for Formula 1 questions. Most of the work was not the model. It was real auth, CI/CD, and Pydantic-typed state. It answers over what it retrieves.

LangGraph · Pinecone · Tavily · FastAPI · Pydantic

chatformula1.com ↗

Assay (FinResearch AI) live

A LangGraph multi-agent research system: a manager fans out to four specialists in parallel, then an analyst and a reporter write a verdict over what came back, all in structured JSON.

LangGraph · Pinecone · RAG · Python

Hugging Face Space ↗

TravelOS live · personal tool

The operations console for my own trips. One user, me, most days. It exists because spreadsheets kept losing arguments with airline schedule changes.

personal ops console

travel.prateekmulye.dev ↗

slipstream-f1-strategist in development

An event-driven race-strategy service in Java, built for correctness under load: a reactive API publishes to Kafka, a consumer runs a deterministic simulation, and the result is written back exactly once.

Java 21 · Spring WebFlux · Kafka · PostgreSQL · OpenTelemetry · Testcontainers

github.com/prateekmulye ↗

The maker

Off the clock.

Prateek at the Koko Head summit on Oahu, ocean behind him. — Koko Head, Oahu

Prateek on the Chicago riverwalk. — The riverwalk, Chicago

The photos are from my own trips. Time off mostly looks like that: I hike when I travel, and I walk new cities instead of driving through them. I follow Formula 1 closely enough that two of the projects above exist because of it.

One habit that shaped how I work: a database once quietly filled its disk overnight and took the app down with it. I recovered it by hand, then added the retention policy, the cleanup job, and the alert that should have been there all along. Disk and observability are the first things I check now, on every system I touch.

Education: M.S. Computer Science, University of Illinois at Springfield (2018) · Post-Graduate Diploma, Advanced Computing, C-DAC, Pune (2013) · Bachelor of Computer Application (BCA), University of Pune (2012)
Certification: Generative AI with Large Language Models, DeepLearning.AI and AWS (Coursera), Feb 2026
Languages: English, Hindi and Marathi. Some Italian, and German at A2 and climbing.

Career

11+ years, in order.

2026 to present

Agilent Technologies

Software Engineer, Manufacturing (Backend & Automation)

I build and fix the manufacturing processes and recipes a production plant runs on. Separately, I build internal applications that put AI to practical use, and I lead the team's internal AI enablement. Internal work, described by outcome.

2021 to 2025

HG Insights

Software Engineer

Standardized 15M+ US and Canada records into canonical geographic shapes, adopted as the framework for later data-ingestion workflows, and cut the heaviest dashboard's latency by more than 60% on a 30M+ row PostgreSQL dataset via partitioning, targeted indexing, and read-replica routing.

2018 to 2021

Capital One, via Cognizant

Software Engineer

Hardened a financial-transaction event platform against message loss and double-processing: a dead-letter-queue strategy, bounded retry with backoff, and idempotent replay. Also unblocked performance testing at 12K+ requests per second with OpenResty service virtualization standing in for unavailable downstreams.

2017 to 2018

Illinois Department of Public Health

Software Engineer Intern

Built a reporting application for state health records while finishing the M.S. on a state-sponsored scholarship.

2015 to 2018

University of Illinois at Springfield

M.S. Computer Science

Between AurionPro and the internship above, I moved from Pune to the United States for graduate school, and finished the M.S. in Computer Science in 2018. The IDPH role was part of this chapter, on a state-sponsored scholarship.

2013 to 2015

AurionPro Solutions

Software Engineer

Built the secure login and anti-phishing layer for a bank's first online-banking platform, in Java, under real regulatory and performance constraints. New from scratch, and it had to be right the first time.

Contact

If you write, I answer.

Ask about a role, a system on this page, or anything the resume leaves open. You will hear back, even when the answer is no.

This record exists in two other forms, because I enjoy building the same thing more than one way. Deep Field draws it in 130,000 particles, rendered live on your GPU. The Record is the long form, with a chat grounded in this site's own text that tells you when it does not know. The switch in the corner moves between them.

Astro, with Newsreader and JetBrains Mono self-hosted, rendered on Cloudflare. Built by hand and iterated with Claude Code. The chat on the Record page runs retrieval over this site's own text on Workers AI. The facts here match the resume. Last updated July 2026.