Umang Dubey Backend AI Engineer
Open to new roles Get in touch
V3.0 / 2026.09.15 Portfolio / Vol. I

UMANG DUBEY

Distributed systems engineer by foundation. I build high-throughput backend platforms and the AI agent infrastructure that operates them.

01

BACKEND
ai engineer.

02
About 01 / A little about me

I'm a Backend AI Engineer building scalable systems where reliability, performance and cost matter.

At CoinSwitch I work on the infrastructure powering a high-volume trading platform that handles 30M+ orders a day, from access control and delegated trading to the systems behind the order pipeline. I've also driven improvements like a 93% reduction in observability cost, using the same approach I bring to performance work: understand the real bottleneck, then engineer around it.

I enjoy taking problems from 0→1: starting with the use case, designing the right architecture, and turning an idea into a production system.

My interests sit at the intersection of distributed systems and agentic AI, from partitioning, backpressure and consistency trade-offs to tool design, orchestration and guardrails.

Always building. Always learning. Looking for the next opportunity.

Experience 02 / Three roles, 2022–2026
Fintech · AdTech · DevTools
Nov 2025 — Present / Bengaluru, India / Fintech · Crypto trading

Senior Software
Engineer

CoinSwitch
1M+ active users

I own access control and reliability for the trading services at the centre of the platform — the path more than a million users trade through every day — and I build the agent tooling the on-call team runs on.

  • Built a delegated, role-based access system as a single unified service, letting 1M+ users securely grant control of their trading accounts to someone else. It is now the authorisation path for 10+ product lines, with every access decision routed through one auditable point.
  • Re-architected the high-volume trading services for scale: modularised the core, enforced idempotent order handling, expanded automated coverage, and took ownership of incident response, cutting MTTR by 60% and taking P95 order latency down by an order of magnitude.
  • Delivered org-wide INR→USDT conversion with real-time rate adjustment and hedging-backed position protection, now processing 30M+ orders daily under strict consistency requirements.
  • Shipped MCP-based AI agents for on-call and operational workflows, and rebuilt logging, metrics and alerting around genuine paging signals rather than volume, cutting observability cost by 93%.
Oct 2024 — Oct 2025 / Bengaluru, India / AdTech · AI marketing platform

Senior Software
Engineer

Pixis.ai
High-concurrency pipelines

Owned performance and concurrency across the core backend, and mentored engineers through design to delivery.

  • Cut memory footprint by 35% and P95 latency by 70% across critical services by eliminating storage inefficiencies, refactoring data models, and migrating graph-heavy workloads to TimescaleDB.
  • Redesigned a high-concurrency consumer service around a bounded worker pool. Unbounded goroutine growth was the real failure mode; tuned concurrency limits and proper backpressure lifted throughput 50%, to 20K+ messages/sec under burst traffic.
  • Built live performance-insight and notification features in Prism, surfacing campaign alerts that let marketing teams react 3x faster.
  • Mentored 3 engineers on code quality and review practice, halving PR turnaround, while owning features from HLD through delivery.
Jul 2022 — Sep 2024 / Noida, India / DevTools · Cloud testing

Software
Engineer

LambdaTest
Infrastructure & platform

Consolidated fragmented infrastructure into single services, and replaced manual, engineer-hours-heavy processes with self-service tooling.

  • Led the consolidation of distributed inventory services into one microservice, cutting user launch time by 45% and eliminating device unavailability through tighter state consistency and service orchestration.
  • Replaced manual database entries for device onboarding with a self-service integration workflow, taking the deployment cycle from a week to a day.
  • Designed an AI-powered accessibility platform end to end, with proof-of-concept integrations for enterprise clients, enabling systematic compliance checks across 15+ services and cutting manual audit effort by 60%.
  • Decoupled the frontend into independently deployable modules, cutting page load time by 40% and cross-team release cycles by 30% across 5+ internal teams.
Systems 03 / Production & self-directed
TRIAGE↓ / Alert-to-remediation loop / Agentic AI · SRE

Self-healing
on-call agent

Claude Agent SDK MCP Prometheus Kafka Kubernetes

An autonomous incident-response system for a microservice stack. Most on-call incidents repeat — a crash-looping worker, a bad deploy, a poison message, an OOM — and each has a known diagnosis path. The slow part is a human waking up and orienting.

Five agents run as a pipeline behind a deterministic Python state machine: detector, diagnostician, remediator, verifier, librarian. Telling failures apart needs reasoning — a stopped consumer and a crash-looping consumer look identical in the lag metric but need opposite fixes — so the agents own judgement while the orchestrator owns lifecycle and sequencing.

The safety layer is code, never the model: an allow-list, a confidence gate and a circuit breaker sit between any proposal and production, and the MCP tool surface separates read-only tools from destructive ones outright.

github.com/dev-umang-here/ai-sre-agent ↗
ORCHESTRATOR deterministic state machine · owns the incident lifecycle DETECTOR alert + signal intake DIAGNOSTICIAN root-cause hypothesis REMEDIATOR restart, scale, rollback VERIFIER did it actually fix? LIBRARIAN runbook store GUARDRAIL LAYER ALLOW-LIST · CONFIDENCE GATE · CIRCUIT BREAKER — IN CODE, NEVER THE LLM MCP TOOLS — READ-ONLY metrics · logs · traces · deploy history MCP TOOLS — DESTRUCTIVE docker · k8s · kafka offsets · rollback OUTCOMES FEED THE RUNBOOK STORE — THE NEXT INCIDENT STARTS WARMER
Fig. 01 — Agents reason. Code decides what they're allowed to touch.
IN PRODUCTION / intellr.io / Multi-agent platform

Intellr
hiring platform

Multi-agent Voice AI Orchestration Event pipelines

A live AI hiring platform built around three autonomous agents behind a shared orchestration layer. A screening agent scores candidates across resume, LinkedIn and GitHub signals; a voice agent runs real-time conversational interviews with adaptive follow-ups; an automation agent advances candidates through rule-driven pipelines.

The engineering problem is not the models, it's everything around them: keeping voice turn-taking inside a latency budget, making scoring reproducible enough to audit, and holding pipeline state consistent while three agents and a calendar integration all write to it. The orchestration layer is deterministic on purpose, so agent behaviour stays observable and replayable.

intellr.io ↗
APPLICANTS resume · profile · repo SCREENING AGENT multi-signal scoring VOICE AGENT adaptive interview AUTOMATION IF/THEN advance ATS + CALENDAR DETERMINISTIC ORCHESTRATION LAYER pipeline state · retries · audit trail EVERY AGENT DECISION IS SCORED, LOGGED AND REPLAYABLE
Fig. 02 — Agents do the judgement. The orchestrator owns the state.
PLANNING↑ / Preferences-to-basket / Agent orchestration

Instamart
grocery agent

MCP tool server LLM planning

An agent built on Swiggy Builders Club's Instamart MCP server that turns dietary preferences into a weekly meal plan, then resolves that plan into a real basket — SKU matching, substitutions when stock fails — and places the order.

Built first for people in Bengaluru who don't know what to cook. Checkout is deliberately approval-gated: an agent that spends your money unprompted is a bug, not a feature.

PREFERENCES diet · budget · household MEAL PLAN LLM planning · 7 days RESOLVE BASKET Instamart MCP server SKU MATCH · SUBSTITUTIONS HUMAN APPROVAL PLACE ORDER gated checkout
Fig. 03 — Planning is automated. Spending is not.
Engineering 04 / How I build and operate systems

01

Architecture

HLD to LLD. I design the system around what it has to solve, and review the designs that ship alongside mine.

02

Ownership

Design through on-call. The rollout, the dashboards, the SLO, and the escalation path when it breaks.

03

Reliability

Idempotency, backpressure, circuit breakers, graceful degradation. Every dependency has a defined failure behaviour.

04

AI-Native

Agents, MCP tool servers and LLM workflows built to production standards: gated, observable, cost-controlled.