Backend & AI systems that survive production.
I'm Arnav, a software engineer in Mumbai. At Sure Financial I build the multi-agent system behind an AI financial advisor that handles 4,000+ conversations a month. Before that I was the only engineer on inboundSelect, building its React app, real-time call routing and Stripe billing.
Fig. 1One request through the Sure Financial advisor: route, tool call, grounding check, reply.
Stack
4,000+
AI conversations a month
Sure Financial
1,000+
monthly active users
Sure Financial
380+
REST endpoints shipped
inboundSelect
3
products live in production
Sure, inboundSelect, Royal Algae
§ 01About
An engineer who cares what happens after deploy.
I like the parts of a system most people hope never break: concurrency, money movement, retries, and getting LLMs to stick to the facts.
For the past year and a half I've shipped production systems end to end, from schema design to deploys. Today I build agentic AI at Sure Financial. Before that I owned the entire backend of a US telephony and billing platform.
B.Tech, Computer Engineering, K J Somaiya College of Engineering, Mumbai (2022 – 2026).
01
Correct under concurrency
Exactly-once call assignment, idempotent payments, isolated async sessions.
02
Grounded AI, not vibes
Agents answer from the user's real data, and every reply is checked before it ships.
03
Observable by default
Tracing, cost per agent and routing-drift detection, wired in from day one.
04
Owned end to end
Schema, API, queues and deploys on GCP, AWS, Railway and Vercel.
§ 02Experience
Where I've shipped.
An AI advisor in production, a payments-and-telephony backend I owned alone, and the APIs that came before them.
- SF
AI Systems Engineer
Full-timeJan 2026 – PresentI build the multi-agent system behind Sure's AI financial advisor, which handles 4,000+ conversations a month from 1,000+ monthly active users.
- Designed a coordinator → specialist multi-agent system on Google ADK and Gemini (advice, in-app actions, loan negotiation), with a parallel “thinking” agent for faster perceived responses.
- Reduced hallucinations with deterministic grounding: compact per-user context cards, plus a post-reply check that every institution the model names exists in the user's accounts.
- Eliminated cross-user context leakage under concurrent async load by isolating request state with Python ContextVars.
- Built a Go (Gin, GORM) microservice that syncs user profiles and events to MoEngage through a governed event catalog.
- Instrumented every agent with Langfuse: tracing, per-agent cost attribution and routing-drift detection.
PythonFastAPIGoogle ADKGeminiGoPostgreSQLFirestoreGCP Cloud RunLangfuse - IS
Full Stack Engineer
inboundSelect· Remote · US client
ContractNov 2025 – Jul 2026The only engineer on a multi-tenant pay-per-call platform that routes live phone calls to licensed insurance agents, and bills for every one of them. I worked directly with the client and built the whole product, from the React frontend to PostgreSQL.
- Built the React 19 web app (146 components): real-time agent and admin dashboards over Socket.IO, TanStack Query, React Router, Stripe's embedded checkout for wallet top-ups and subscriptions, and Recharts analytics.
- Built real-time call routing on Twilio and Socket.IO with exactly-once agent assignment, even when two agents accept at the same instant.
- Shipped multi-tenant payments on Stripe Connect: prepaid wallets, prorated subscriptions and per-call fees, idempotent end to end.
- Automated invoicing, fee collection and recovery with 9 scheduled Bull/Redis jobs that retry with backoff and survive a Redis outage.
- Designed 380+ REST endpoints over a ~60-table PostgreSQL schema (105 migrations).
ReactTanStack QueryNode.jsExpressPostgreSQLRedisBullSocket.IOTwilioStripe ConnectAWSRead the case study - EX
Backend Engineer Intern
Exploria
InternshipApr 2025 – Jun 2025Built REST APIs and AWS-backed workflows for the product's backend.
- Built REST APIs in Node.js and Express with request validation, tests and API documentation.
- Integrated AWS S3 storage and AWS Lambda functions into backend workflows.
Node.jsExpressAWS S3AWS Lambda
§ 03Selected work
Systems I've built that are live.
Every one of these is in production. Open a case study for the architecture and the decisions behind it.
- AMlicensed FL · available
- RKlicensed FL · available
- JTlicensed TX
- SPon another call
UPDATE calls SET agent_id = $1 WHERE id = $2 AND agent_id IS NULL
→ wallet debited once · idempotency key ok · stripe connect
inboundSelect
Real-time call routing and billing for US insurance agencies
A multi-tenant pay-per-call platform. Live calls reach exactly one licensed, available agent within seconds, and every connected call is billed through Stripe Connect. As the only engineer, I built the whole product: the React app, the backend and the database.
- REST endpoints
- 380+
- Postgres tables
- ~60
- scheduled jobs
- 9
- verified webhooks
- 9
Royal Algae
A D2C store for a sea moss wellness brand
A live e-commerce store with phone-OTP guest checkout, Razorpay and cash-on-delivery, server-side GST pricing and PDF invoices. I built the React storefront and admin, and most of the checkout backend.
- API routes
- 48
- PHPUnit tests
- 13
- payment modes
- 2
- in production
- Live
- area
- 42,000 sq ft
- rent / sq ft
- ₹28 ₹31
edited
in review
approved
employee rent ₹28 → ₹31
pending PendingAction → manager
approved change applied to record
BuildFlowCRM
A CRM for commercial real estate with maker-checker approvals
A MERN CRM where five property types share one schema family, and every employee edit or delete waits for a manager's approval before it lands.
- property types
- 5
- access control
- 2-tier
- import & export
- Excel
- API tests
- Vitest
§ 04Stack
The toolbox.
What I reach for, grouped by where it sits in a system.
AI & agents
Multi-agent orchestration, grounding and LLM observability.
Backend
APIs, real-time systems and background work.
Data
Relational first, with caches and queues around it.
Cloud & DevOps
Shipping and running it.
Payments & comms
Money movement and telephony.
Frontend
Enough to ship the product around the API.
§ 05Open source
Write the schema once.
Package
mongoose-zod-schema-builder
Define a schema once and get both a Mongoose model schema and a Zod validator from it, so what your database stores and what your API accepts can't drift apart.
- npm downloads
- 400+
- releases
- 3
npm i mongoose-zod-schema-builder1import { createSchemas } from "mongoose-zod-schema-builder";2 3const user = {4 name: { type: "String", required: true, minLength: 3 },5 email: { type: "String", unique: true, email: true },6 password: { type: "String", minLength: 8 },7};8 9// one definition → a Mongoose schema and a Zod validator10const { mongooseSchema, zodSchema } = createSchemas(user);§ 06AI assistant
Skip the scrolling. Ask my AI.
An assistant that knows my work in detail. Ask it anything you'd ask me on a first call.
- 01Grounded on my real experience and projects, and told never to invent facts.
- 02Gemini with function calling, streaming replies as it writes.
- 03Can pass your name and message straight to my inbox.
- 04Per-visitor token quotas and a monthly spend cap keep it affordable.
Arnav's assistant
● online
Try asking
§ 07Contact
Let's build something that lasts.
I'm open to full-time SDE, backend and AI engineering roles. Email is the fastest way to reach me, or send a note here and it lands straight in my inbox.
- arnav@arnavlogic.com
- arnav-kumar
- GitHub
- arnav14-dev
- Resume
- Based in
- Mumbai, India · IST