Building AI agents at Reliance Intelligence · open to conversations

Backend systems and AI agents that hold up at scale.

Backend & AI engineer in Bengaluru with 3+ years taking products from 0 to 1. Core engineer behind CRED Garage and Insurance (7M+ users), now building agents and skills for Bharat AI.

Email me Résumé LinkedIn GitHub
7M+users on products I built
₹350Cr+annual revenue across them
~1000requests/sec at peak
~1Mdaily active users

Systems

Things I've designed and shipped

Simplified sketches of the architecture behind my work and side projects. Hover a box to trace what it talks to.

Press play to follow one request through this system, step by step, with narration.

Simplified: general patterns I used, not proprietary designs. Swipe the diagram sideways to see all of it.

Experience

Where I've worked

2019
IIT JodhpurB.Tech CSE · CGPA 8.22
2022
SMS Datatech, TokyoIntern → pre-placement offer
2023
CREDGarage & Insurance, 0 → 7M+ users
2026
Reliance IntelligenceIntelligence platform, chat + voice
Sep 2026 – now

Backend AI Engineer · Reliance Intelligence Current

Building the intelligence platform behind Bharat AI, which makes AI accessible to Indian users at scale.

  • Onboarded cricket, music and news as capabilities on the platform's orchestrator agent, giving users real-time scores, playback and headlines in natural language.
  • Worked across the voice stack (speech-to-text, text-to-speech and speech-to-speech) and built services in a hexagonal (ports and adapters) architecture, with MLflow for model tracking.
  • Brought music to speech-to-speech (S2S) voice calls and handled barge-in: when a caller speaks over a playing song, playback stops and the agent responds.
  • Built a Watcher Service that monitors configurable events and proactively sends users contextual reminders, running on a workflow service built on Temporal for durable schedules and retries.
PythonAI agentsLLMsS2SSTTTTSMLflowTemporalHexagonal architectureMicroservices
May 2023 – Aug 2026

Backend Engineer · CRED

Built and scaled CRED Garage and Insurance from 0 to 1.

  • Grew the products to 7M+ users and ₹350Cr+ in annual revenue; ran services at ~1000 RPS and ~1M DAU with steady latency through traffic spikes.
  • Architected a cross-sellable pocket insurance platform across multiple lines of business: ~₹40Cr revenue.
  • Built a motor insurance assistance flow that lifted conversion by 13%, and a pluggable self-inspection platform that raised quote coverage by 11% with zero-effort insurer onboarding.
  • Hardened high-throughput insurance flows with circuit breakers, idempotency keys, retries with DLQs, Sagas and event-driven design on Kafka; distributed rate limiting on Redis; JVM and query tuning.
  • Shipped Garage features: challan details and payment, FASTag recharge and passbook, PUCC, a scraping platform, and car valuation with CARS24 and Spinny.
  • Shipped an on-call AI agent (OpenAI + RAG) that diagnoses incidents through MCP-exposed microservices, and an AI layer for KYC review that cut ops work by 30%.
JavaSpring BootPythonKafkaRedisMySQLDynamoDBElasticsearchgRPCAWS
May – Jul 2022

Software Engineer Intern · SMS Datatech, Tokyo Pre-placement offer

  • Integrated 10+ SaaS apps into a unified employee-management platform through a pluggable connector framework (Strategy and Factory patterns): no core changes per new integration.
  • Built a PyTorch activity classifier that flags anomalous SaaS usage.
JavaSpring BootPyTorch

Projects

Things I build for myself

Side projects, mostly AI agents and developer tools. Several are open source on GitHub.

Coding agent · open source

kmare-code

A coding-agent harness built from scratch, a smaller cousin of Claude Code.

The model is the brain; the harness is the body. A planner breaks the task down, a generator does the work with tools (read, write, run a command), and an independent evaluator checks the result before anything is called done. Point it at your test command and a failing run fails the round straight away.

  • kmare-chatstreaming chat UI with prompt, exact-match and semantic caching
  • kmare-voiceself-hosted push-to-talk: faster-whisper STT, Piper TTS
  • kmare-observabilitymetrics and traces for every LLM call
6model providers, plus an offline mock
3independent roles: plan, build, check
PythonAgentsTool useFastAPISQLiteCaching
View on GitHub ↗
kmare-code / harness-engineering.html
The harness-engineering guide in the repo
Voice assistant · open source

Escanor

A local, Jarvis-style assistant for the Mac, by text or voice.

One async agent sits behind a terminal, a web HUD and a hands-free voice loop. It keeps a second brain of notes with keyword and semantic search, runs reminders and calendar, plays YouTube Music as an AI DJ, teaches research papers, and plugs in any MCP server as a tool group. A permission gate asks before anything acts outside the chat, and web text is fenced as untrusted.

  • Voicelocal Whisper STT and Piper TTS, replies spoken sentence by sentence
  • Traceswhere every turn's seconds went: model, tools, speech
  • ModelsClaude, GPT, Gemini or a fully local Ollama model
359offline tests
10/10correct tool choices on the Gemini eval
PythonVoiceMCPRAGOllamaPrompt-injection safety
View on GitHub ↗
localhost:8770
The Escanor HUD: hold Space to talk
Open source library

mayank-awesome-skills

31 reusable Claude Code skills, curated and made generic.

Engineering-stack skills for Go, Java, Node and Python microservices, Flutter, React and AI agents, each with references and starter templates. Workflow skills for TDD, implementing a feature from a spec, autoresearch loops, technical design docs, QA and saving session state. The memory-coach skill ships its own interactive workbook site.

  • Stack skillsframework choice, persistence, caching, messaging, observability
  • Workflow skillstdd, implement-feature, autoresearch, qa, brace-for-compact
  • Used byEscanor loads its tutor skill from this library
31skills
7engineering stacks covered
Claude CodeAgent skillsMarkdownHTML
View on GitHub ↗
mayank9mare.github.io/mayank-awesome-skills
The Memory Palace Workbook, published with GitHub Pages

More projects

AI agentPrivate

Job Apply Agent

Finds openings at companies I choose, picks the best résumé version, fills applications in a real browser and submits only after I approve by email. Built test-first, with 325 tests.

PythonFastAPIPlaywrightGeminiSQLite
AI media editorPrivate

EditorMare

Local-first photo and video editing: turns a written brief into a reviewable plan, runs approved steps through MCP, verifies the result and recovers from failures. Cuts Shorts and Reels from a finished video.

PythonMCPCircuit breakersTracing
Trading engineGitHub ↗

TradeMare

A swing-trading engine for Indian and US equities: fundamental, technical and LLM sentiment signals, budget and risk caps, and no order without my approval. Early development.

PythonDockerLLM sentimentStrategy pattern
Backend serviceGitHub ↗

Job Scheduler

A cron-style job scheduler service: register a job with a cron expression, have it fire on time, and query its run history: status, duration, exit code and output.

Java 21Spring Boot 3.5MySQLJPA
macOS appPrivate

ClaudePlane

A floating macOS panel that tracks every Claude Code and Codex session across VS Code, IntelliJ and Terminal, and jumps to any of them with one click. Highlight part of the screen to ask any model about it.

SwiftmacOSHooks
Slack botGitHub ↗

Guts

Drive Claude Code terminal sessions from your phone through Slack, over a Socket Mode WebSocket to a server on your Mac.

PythonSlackasyncio

Research & earlier work

NEFD: Noise Engineered Proxyless Federated Distillation

A federated learning framework that handles model heterogeneity without proxy datasets. Beat state-of-the-art baselines; accepted at IEEE WCCI 2024.

Federated learningDistillationPyTorch

UPI Aggregator

One app to compare cashback offers across UPI providers, with a modular plugin architecture.

Spring BootAndroidPlugins

Toolbox

What I work with

Languages

JavaPythonC++JavaScriptSQL

Backend

Spring BootDjangoRESTgRPCMicroservices

Infra & data

KafkaTemporalRedisAWS SQS · S3ElasticsearchMySQLPostgreSQLDynamoDB

Architecture

System designHexagonal (ports & adapters)Event-drivenSagasCircuit breakersRate limiting

AI

AI agentsRAGMCPLLM integrationSpeech: S2S · STT · TTSMLflowPyTorch

Tools

GitClaude CodeCursorPostman
#34ICPC Asia Regionals
99.5percentile, JEE Main & Advanced
8.22CGPA, IIT Jodhpur CSE
WCCIIEEE 2024 publication
Codeforces · Kmare ↗ LeetCode · Knightmaree ↗ CodeChef · K_mare ↗

Writing

Notes from building things

Short write-ups on patterns I've used in production and in side projects. Each one has a Listen button if you'd rather hear it.

Workflows · 3 min read

Durable reminders with Temporal

How a “tell me when the match starts” feature becomes one small, crash-proof workflow instead of a cron job, a table and a pile of retry code.

Read or listen Distributed systems · 3 min read

Issuing a policy without losing a step

Payment succeeded, the insurer's API timed out. How a chain of small, idempotent queue steps keeps a purchase from getting stuck or happening twice.

Read or listen AI agents · 2 min read

A model can’t grade its own homework

What I learned building kmare-code, a coding-agent harness from scratch: the model is the brain, but the harness decides whether the work is actually done.

Read or listen

Contact

Building something that has to scale? Let's talk.

Open to senior backend and AI engineering roles.

mayank42293@gmail.com LinkedIn Résumé (PDF)