I build back-end systems for the case where things go wrong. 11+ years building distributed systems where a dropped or double-processed message is real money moving the wrong way, and PostgreSQL kept fast past 30M rows. Now I ship agentic AI, RAG and multi-agent systems, with the same discipline: they answer from what they retrieved instead of guessing.
Same engineer, same facts. This is the system I am building next. Roll back anytime.
Open to the right senior role, remote or relocation.
Scroll for the details, or jump straight to the resume. Press Cmd K to jump anywhere.
ChatFormula1 Agentic RAG for Formula 1 questions. Single agent, vector store plus live search, Pydantic-typed state. open the live demo → 
FinResearch AI LangGraph multi-agent research: a manager fans out to specialists; every step emits structured JSON. open the live demo → 
slipstream-f1-strategist Event-driven race strategy in Java: WebFlux, Kafka, Postgres. Bounded retry, idempotent writes. read the code → The chat is one router and one retrieval pass. FinResearch AI, the multi-agent system, is a project this chat describes, not a capability this page runs.
The work below is the proof: a payments queue where a bad message can't slip through, a 30M-row database I kept fast as it grew, and a multi-agent system that answers only from what it retrieved.
Hardening that platform against message loss and double-processing was my job: a dead-letter-queue strategy, bounded retry with backoff, and idempotent replay, so a re-run never double-charges anyone. I watched backlog per partition above everything else, because one poison message sits at the head of a partition and stalls every message behind it.
The diagram above is that path: one poison message, set aside and replayed exactly once, nothing lost or duplicated. A dead-letter queue with bounded retry and idempotent replay, the pattern from a payments platform.
Past 30M rows, a PostgreSQL database stops being forgiving. I cut the heaviest dashboard's latency by more than 60% with range partitioning by region, targeted indexes around the queries people actually ran, and read-replica routing that kept the heavy reporting reads off the live path.
People spell the same place a dozen ways, and the reports built 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.
On the same platform I cut contract-processing turnaround by about 60% across an 80K+ record pipeline: idempotent Oban Pro workflows with bounded retries and exponential backoff, in Elixir and Phoenix. I raised company-spend freshness from a monthly batch to a weekly cadence with scheduled, on-demand, and event-triggered jobs. And I made our reliability targets enforceable: end-to-end OpenTelemetry, Datadog SLOs, and reversible releases with Argo CD and LaunchDarkly. I mentored the mid-level engineers who later carried it.
Structured output and grounding are the AI version of a queue that won't drop or duplicate a message: don't let the system make things up or lose what it had. FinResearch AI is a multi-agent system where 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.
The chat on this page is one of those builds: it answers from this site's own content, runs on Cloudflare Workers AI, and when it doesn't know something it tells you and points you to email.
Open FinResearch AI on Hugging Face →11+ years keeping distributed systems correct. Now I build systems that reason.
a zero-downtime deploy of this page. Roll back anytime.
I read every message myself and I reply. Open to the right senior role, remote or relocation.
There is a second side to this page. · roll back anytime.
You are on the Console. whenever you like.
A couple of things about how I work, and what I build when it's just me.
Everything as plain text. Senior Full-Stack Engineer, backend-heavy, 11+ years of professional IT experience, distributed systems with a real, growing practice in applied AI.
Download my resume (PDF) ↓An event-driven race-strategy service in Java, built for correctness under load. A reactive API publishes a request to Kafka, a consumer runs a deterministic simulation, and the result is written back exactly once. Same shape as the financial-grade reliability work below, minus the NDA: bounded retry with backoff and idempotent processing, so a bad message gets retried sensibly and is never processed twice.
read the code →An agentic RAG assistant for Formula 1 questions. The model is the easy part; the work was the production hygiene around it, so it has real auth, CI/CD, and Pydantic-typed state instead of being a notebook demo. A single agent routes between a vector store and live web search, then answers over what it retrieves.
open the live demo →A LangGraph multi-agent system I architected and deployed. A manager fans out to specialized research agents that run in parallel, then an analyst and a reporter write a verdict over what came back. Every step has to emit structured JSON outputs, which keeps the agents grounded instead of inventing things. I built it as my submission to the SuperDataScience CP044 community project.
open the live demo →A reference architecture for watching an AI agent's tool calls. It sees each call, runs an advisory policy gate and a tamper-evident audit, and does all of it out-of-band, never in the agent's hot path. It observes and records; it never decides what runs. Written in Java, the stack I reach for when uptime is the requirement.
An early build that takes on the tax and visa side of moving between countries, a problem I'm living through myself right now. Still pre-release.
Roles by public label, no client names. 11+ years across financial, banking, and data-intelligence systems.
I design ingestion pipelines for production telemetry, the data foundation for analytics and AI, and harden the backend logic of on-premise production systems for fault tolerance on constrained hardware. I also build and fix the processes and recipes the plant runs on, and lead internal AI enablement. Internal work, described by outcome.
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.
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.
Built a reporting application for state health records while finishing the M.S. on a state-sponsored scholarship.
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.
DeepLearning.AI and AWS (Coursera) · verify the credential →
University of Illinois at Springfield
C-DAC, Pune
University of Pune
One handle across GitHub, LinkedIn, and Hugging Face, cross-linked, so a reviewer can check any claim here against the live source.
The code behind the claims, public and in the open.
The full work history, employers public by name.
connect on LinkedIn →Assay (FinResearch AI), a multi-agent Space.
open the demo →A RAG assistant for F1 questions.
open the demo →Open to senior roles, remote or relocation · 11+ years of professional IT experience
No matching action.