Open to GenAI and backend roles · Noida or remote

I build AI that doesn't make things up.

On backends that don't fall over. I work in Java and Python with RAG, agent workflows and evaluation, and every number on this page is one I measured.

See the work Download CV

AgentOps · SDLC run · coupon feature for ShopLite illustration

Pipeline

Requirements
Design
Tests first
Build
9 quality gates
Review
Pull request

Workers

sdlc · requirements

sdlc · build (waiting)

invoice workflow

idle · waiting for a lease

steps 0 / 7 gates 0 / 9 side effects repeated 0

Event log

The window shows the flow. Measured on the real engine: 500 runs · 435 random worker kills · 0 lost · 1.9 s median takeover

Measured Nov 2026. Each number links to how it was measured.

Selected work

Things I built, running.

Each card plays a short illustration of what the project does. The case study has the design, the hard parts and the measured results.

chaos test

runs 500/500workers killed 435lost 0

AgentOps

Live demo

Agent workflows that survive crashes, wait for people, and ship code through SDLC gates.

0 lost runs in 500 · 435 random kills

Live demo GitHub

Python · FastAPI · Postgres · Docker · AWS Lambda · Terraform · LLM Video Case study

job description check

Python + FastAPI servicesverified · cited code

RAG with a vector storeverified · cited code

LangGraph agentsverified · cited code

Kubernetes in productionunverified · no code

FitCheck

Live demo

Paste a job description; every requirement is checked against real code, or marked unverified.

100% accuracy · 0% false claims

Live demo GitHub

LangGraph · FastAPI · RAG · pgvector · Evals Case study

claim decision

Claim

7 rules

AI check

Decision

APPROVErules allow, AI agrees

Health Claim Agent

AI-assisted claim decisions where the Java backend, not the model, makes the final call.

122 tests · AI called only after 7 rules pass

Live demosoonGitHubsoon

Java 21 · Spring Boot · pgvector · React Case study

scan

const express = require("express")

const app = express()

const KEY = "sk_live_••••••••"

app.get("/user", (req, res) => {

const q = "SELECT * FROM users WHERE id=" + req.query.id

db.query(q, (e, rows) => res.json(rows))

})

app.post("/login", login)

app.listen(3000)

HIGHhardcoded secret · line 3

HIGHSQL from user input · line 5

MEDIUMno rate limit on login · line 8

gradeF

VibeAudit

Live demo

Finds the bugs that AI-written code ships with, before that code reaches real users.

Every planted bug found in 39.8 s · 2 model calls

Live demo GitHub

FastAPI · LLM · SSE · Docker · bandit · semgrep Case study

How I build

AI is never the last word.

Rules, tests, gates or a person decide. The same three habits run through every project.

01
AI proposes, code decides

In the Health Claim Agent, seven rules run before any model, and every rejection is rule-driven. In FitCheck, proof must be retrieved code, or the requirement is marked unverified.
See the proof
02
Break it on purpose

AgentOps ran 500 workflows while workers were killed at random, 435 times. Nothing was lost and no side effect ran twice. Measured, not assumed.
See the proof
03
Measure, then claim

FitCheck is scored against a keyword baseline on 25 labelled requirements. If a number is not measured yet, it is not on this site.
See the proof

Skills

Every skill points at its proof.

No bars or percentages. Each group links to the projects where you can read the code.

Backend

Java 21
Spring Boot 3
Spring Security (JWT)
JPA
FastAPI
REST
SSE

Health Claim Agent VibeAudit AgentOps

GenAI

RAG
pgvector
Embeddings
Structured outputs
LangGraph
Evals
Grounding

FitCheck Health Claim Agent

Agents and reliability

Durable workflows
Leases and fencing
Idempotency
Human in the loop
Sandboxing

AgentOps

Data

PostgreSQL (SKIP LOCKED, LISTEN/NOTIFY)
pgvector
DynamoDB

AgentOps FitCheck

Cloud and DevOps

Docker
GitHub Actions
AWS Lambda
EventBridge
Terraform
S3 + CloudFront

AgentOps Health Claim Agent

Frontend

React + TypeScript
Vanilla JS UIs with SSE
Astro

Health Claim Agent VibeAudit

About

Where I've worked and learned.

Experience · Feb 2026 – present

GenAI Developer, for a US healthcare client

At TCS · Noida

Build agentic AI for insurance claim workflows: agents that read claim information, retrieve the relevant policy knowledge and recommend a decision with its reasons.
Work with LLMs, RAG, LangChain and LangGraph, and connect AI services to Java and Spring Boot backends with structured, validated output.
Investigate production issues through Java logs, SQL data and backend behaviour, with development and business teams.

Education · 2021 – 2025

B.E. Computer Science

Chandigarh University

Publication

Automated Network Vulnerability Assessment with Nmap: A Comprehensive Approach

IEEE ICACCTech 2024 · first author

Certifications

Claude Certified Architect
Claude Certified Developer Foundations

Contact

Hiring for GenAI or backend? I'd like to talk.

Open to GenAI and backend roles · Noida or remote

Email me Download CV LinkedIn GitHub

or write to arshdeep.si.3010@gmail.com

I build AI that doesn't make things up.

Pipeline

Workers

Event log

Things I built, running.

AgentOps

FitCheck

Health Claim Agent

VibeAudit

AI is never the last word.

AI proposes, code decides

Break it on purpose

Measure, then claim

Every skill points at its proof.

Backend

GenAI

Agents and reliability

Data

Cloud and DevOps

Frontend

Where I've worked and learned.

GenAI Developer, for a US healthcare client

B.E. Computer Science

Automated Network Vulnerability Assessment with Nmap: A Comprehensive Approach ↗

Hiring for GenAI or backend? I'd like to talk.

Automated Network Vulnerability Assessment with Nmap: A Comprehensive Approach