Senior Backend Engineer & AI Practitioner 11+ years
Systems that stay honest when things go wrong.
Every particle on this page is a message in flight. For 11+ years I have built the systems that carry them: payment platforms where a dropped or double-processed message is real money moving the wrong way, data platforms kept fast past 30M rows, and now agentic AI held to the same standard, answer from what you retrieved, or say so.
Scroll. The ink remembers everything it has carried.
Chapter 01 2013 to 2015 · AurionPro Solutions
The first system had to be right the first time.
A bank's first online-banking platform, built new from scratch in Java under real regulatory and performance constraints. My part was the front door: I built the secure login and anti-phishing layer, the piece where a mistake is not a bug ticket but someone's account. No second deploy fixes a breach.
greenfield · regulated · the shield you are looking at is that login layer, drawn from memory
Chapter 02 2018 to 2021 · Capital One, via Cognizant
On a payments queue, one bad message can stall everything behind it.
I hardened a financial-transaction event platform against message loss and double-processing, measured by the thing that matters on payments, a re-run never double-charges anyone, by building a dead-letter-queue strategy, bounded retry with backoff, and idempotent replay. I watched backlog per partition above everything else, because one poison message sits at the head of a partition and holds up every message behind it.
When unavailable downstreams blocked performance testing, I unblocked it at 12K+ requests per second with OpenResty service virtualization standing in for them.
watch the red ink: poison messages divert to the dead-letter pool, retry on a backoff, and replay exactly once, rejoining green. nothing lost, nothing duplicated.
Chapter 03 2021 to 2025 · HG Insights
Fifteen million records, one canonical shape.
People spell the same place a dozen ways, and the reports on top were comparing two spellings of one city. I standardized 15M+ US and Canada company records into canonical geographic shapes, one consistent country, state and city tree, through a geo-standardization service with shadow-writes and rollback, rolled out in phases with the EVP of Data Solutions. It was adopted as the framework the later data-ingestion workflows were built on.
Past 30M rows, PostgreSQL stops being forgiving. I cut the heaviest dashboard's latency by more than 60%, by range-partitioning by region, indexing for the queries people actually ran, and routing heavy reporting reads to a replica, off the live path.
chaos snaps to lattice; the lattice splits into shards. that is the whole story of this chapter.
Chapter 04 2026 to present · Agilent Technologies
Software that a physical plant runs on.
Software Engineer, Manufacturing (Backend & Automation). I build and fix the manufacturing processes and recipes a production plant runs on, build internal applications that put AI to practical use, and lead internal AI enablement, helping the team adopt it well. Internal work, described by outcome; when a recipe is wrong here, the cost is physical.
five stations, one conveyor. the brass one is the recipe being fixed today.
Chapter 05 on my own time · applied AI
New tools. Same discipline.
Structured output and grounding are the AI version of a queue that will not drop or duplicate a message: do not let the system make things up or lose what it had. Assay (FinResearch AI) is a LangGraph multi-agent system where a manager fans out to four specialists in parallel and every step emits structured JSON, so the agents answer from what they actually retrieved. When the evidence is thin, it says so rather than guess.
ChatFormula1 is the same instinct pointed at a hobby: an agentic RAG assistant, live at chatformula1.com, with real auth, CI/CD and Pydantic-typed state instead of a notebook demo.
one manager, four specialists, and the retrieval shell behind them. the constellation is the actual graph.
Chapter 06 running right now
The living organs.
Claims are cheap. These are deployed systems you can open in another tab right now, each one an organ of the same body of work.
Complaint intelligence over the public CFPB consumer complaint database: it turns vague free-text complaints into standardized, evidence-backed failure modes. Designed grounded or silent, so a classification cites its evidence spans or abstains, and low-confidence cases route to a human.
An agentic RAG assistant for Formula 1 questions. The model is the easy part; the work was the production hygiene around it: real auth, CI/CD, Pydantic-typed state. One agent routes between a vector store and live web search, then answers over what it retrieves.
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. Every step emits structured JSON, which keeps the agents grounded instead of inventing things.
The operations console I run for my own trips. Built to be used daily, not demoed. It is the same instinct as everything above: when the data of your life gets messy, build the system that keeps it honest.
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. Bounded retry with backoff; idempotent writes.
Chapter 07 steady state
If you write, I answer.
I am open to senior backend and applied-AI roles, remote or relocation. The address below goes straight to me. Ask about a role, a project, or anything on this page, and you will hear back, even when the answer is no.
- email[email protected]
- resumeResume (PDF)
- the recordThe long-form record: the chat, the resume spine, the details
- linkedinin/prateekmulye
- githubgithub.com/prateekmulye
- huggingfacehuggingface.co/prateekmulye
The name is borrowed from the Hubble Deep Field: point a telescope at a dark, empty patch of sky for long enough and it turns out to be full of structure. Careers are like that too. Set in Newsreader and JetBrains Mono, self-hosted. Built with Astro, rendered on Cloudflare. The ink is a three.js scene of up to 130,000 particles drawn by your GPU, the sky behind it is a shader, and the sound is synthesized live in your browser, no recording, and off until you ask. All of it is decoration: with JavaScript off, reduced motion on, or a phone in hand, every word on this page is still here. The long-form record, with a chat that answers from these same facts, lives at /record.