AI engineer who ships end to end

Hi 👋, I'm

Rishabh Sharma

Rishabh Sharma
500+ employees get routine HR answers same-day
82% of 5,500 documents a month auto-accepted
85%+ of HR agent tasks passed in evals
3 B2B platforms owned end to end at Zeonix

About how I got here, and what I am good at

I take AI features from the first request to production: the model work, the backend, the interface, and the evals that show whether it works. The model reads; code or a person makes the call.

AI Engineer at Kleeto · 4 AI systems shipped since May 2026, 3 in production

From request to rollout

The HR agent went from requirements to a rollout for 500+ employees. I built Kleeto AI Slides alone and reshaped its authoring flow on its owner’s feedback; 3 customers are piloting it. At Zeonix I shipped 3 platforms, from API design to client onboarding.

Measured before I trust it

The HR agent is tested on known tasks and 50+ attempts to trick it; none worked. On my own help desk, the AI makes the right call on 93.6% of 600 tickets, and an AI grader that disagreed with people was replaced by human labels.

The model reads, code decides

GPT-4o only reads a document’s facts; 9 code rules decide and tell the candidate what to fix, and a model outage falls back to the old flow. The HR agent can never see more than the person asking, and sensitive actions wait for approval.

The longer version

I started in security and moved into AI through the work, not a course. I spent two years at Zeonix on payment and onboarding systems, and shipped my first production LLM feature there: a document-authenticity check on Gemini, written as fixed LangChain chains with retrieval grounding so a human verifier could see why a document was flagged. The model advised and a person decided, and that is still how I build: at Kleeto the model reads, and code or a person makes the call. I care most about what a model can reach, whose permissions it borrows, and whether I can reproduce every number I publish from a clean clone.

Experience most recent first

  1. May 2026 to Present

    Kleeto

    AI Engineer·Gurugram
    May 2026 to Present

    I own Kleeto’s AI layer: 4 AI systems shipped since May 2026, three of them in production for 500+ employees and 2 enterprise clients, and one piloting with 3 customers.

    HR assistantAI agent across the HR platform · 500+ employees
    Document verificationGPT-4o checks on onboarding documents · 5,500 a month
    TheMaskMasks ID numbers before archiving · 30,000 documents
    Kleeto AI SlidesReal-time voice AI presenter · piloting with 3 customers
    • Python
    • FastAPI
    • LangGraph
    • MCP
    • RAG
    • Claude
    • OpenAI
    • Pipecat
    What I did there

    HR assistant

    AI agent across the HR platform · 500+ employees

    • Staff ask for things in plain English across 6 parts of the HR platform (hiring, onboarding, attendance, payroll, leaving, and the document system) instead of hunting through screens: 250+ questions a week, answered the same day instead of in one to two.
    • Built in Python and FastAPI with 18 actions the assistant can take, grounded in 500,000 documents. It asks for approval before anything sensitive, and it can never see or do more than the person asking it.
    • Measured, not assumed: 85%+ of tasks completed correctly, no successful attempt to trick it out of 50+ tries, answers at 6 seconds p95 for 2 cents each.

    Document verification

    GPT-4o checks on onboarding documents · 5,500 a month

    • Checks every document a new hire uploads (right type, readable, clean scan, name matches) and tells them exactly what to fix: 5,500 documents a month, 82% accepted automatically, live for 2 enterprise clients since July 2026.
    • The model only reads the facts off the page; plain code makes the decision, so every rejection names the rule and the number. If the model is down, onboarding carries on the old way instead of blocking anyone.

    TheMask

    Masks ID numbers before archiving · 30,000 documents

    • Hides government ID numbers on identity documents before they are archived, so a full server breach reveals nothing. 30,000 documents masked so far.
    • The customer can still recover their number by answering 4 of 6 security questions. The encrypted copy lives inside the masked image itself, so the server keeps no database and no files to leak.

    Kleeto AI Slides

    Real-time voice AI presenter · piloting with 3 customers

    • Presents any slide deck aloud in English or Hindi, stops for questions at any moment, and answers only from the company’s own documents, with a citation on screen, or says plainly that they don’t cover it.
    • Works on whatever laptop and projector a room has: it tells a question from a cough, a side remark, or its own voice coming back through the speakers. 1,000+ tests; 40 sessions run in the pilot so far.
  2. Jun 2024 to Apr 2026

    Zeonix Global Pvt. Ltd.

    Software Developer·Chandigarh
    Jun 2024 to Apr 2026

    Built and shipped three business platforms, owning each one from the API through to security, integrations, deployment, backups and getting clients set up.

    ZeoCRMUniversity management · all 40+ Australian universities
    ZeoVerifyDocument verification & digital onboarding
    ZeoForexFinancial remittance & currency exchange
    • Node.js
    • Express
    • PostgreSQL
    • GraphQL
    • Angular
    • WSO2
    What I did there

    ZeoCRM

    University management · all 40+ Australian universities

    • Replaced manual data entry with a spreadsheet importer that checks and loads 10,000+ records at once, cutting the job from about three hours to under one.
    • Made the busiest pages 40% faster (1.5s → 0.9s) on 10K requests a day, and set up single sign-on with five permission levels: no unauthorised access in 18 months across 500+ users.
    • Cut commission payouts for 100+ agent partners from two days to same-day by automating the chain from accounts to PDF invoice to email, and saved 4–6 hours a week of form filling with a bot that submits university applications.

    ZeoVerify

    Document verification & digital onboarding

    • Cut document checking from about 25 minutes to 10 per application. Instead of accepting an uploaded scan, the system asks permission and fetches the document from the government directly, so there is nothing to forge.
    • Launched the company’s first production AI feature: a document-checking service in Python and FastAPI on Google’s Gemini, kept deliberately simple and predictable so a reviewer could see exactly why it flagged something. The model advised; a person still decided. This is where my AI work started.

    ZeoForex

    Financial remittance & currency exchange

    • Cut mistakes on live money transfers by a third (8% → 5%) with step-by-step checks and tax rules, and by making each transfer either complete fully or not at all.

Projects built, shipped, and live

Focus AI engineering Backend Platform / DevOps Full-stack
Support queue with response clocks Support queue with response clocks
The support queue
A ticket showing the AI's answer A ticket showing the AI's answer
One ticket, and how the AI answered it
Chat widget on another company's page Chat widget on another company's page
The chat box, on someone else's site
Analytics dashboard Analytics dashboard
Analytics, behind its own permission

Nivara Desk

Flagship 2026

A help desk that many companies share, where the AI answers only when it should

Two failures every AI support product has to rule out: answering when it shouldn't, and showing one customer's data to another. This one is built so neither can happen.

  • AI layerPython
  • APITypeScript
  • Front endsTypeScript
Watch it work (45 s) Try it, signed in Chat box Product API reference Source Write-up
The full write-up
The AI answers from its help centre, hands the rest to a person, and every published number is regenerated from committed data by a script a test re-runs.

Someone asks a question from a chat box on their supplier’s website. It becomes a real ticket in the support queue with a response clock running. The AI answers if it is confident; otherwise a person picks it up.

Guarantee One company can never see another’s data; the database enforces it, not the code
Built with Postgres row-level security · a trained answer gate · replayable tests
Size 4 apps · 3 services · 18 permissions · 550 labelled test questions
Runs on Free hosting, and docker compose up with no API keys

Built as 3 separate pieces

AI layer

Python

Anything the model writes on its own can never reach a customer.

  • Made the right call on 93.6% of 600 recorded tickets (595 of 600) by routing every answer through a trained gate that decides when to reply, when to ask, and when to fetch a human. Money questions always go to a human.
  • Put the right document first 94.2% of the time with hybrid retrieval over Qdrant, and every reply is built from one of those real documents, so a made-up answer is impossible rather than unlikely. No attempt to trick it has worked, and the attacks re-run on every change.
  • Held wrongful escalations to 6.8% with zero wrong answers on money questions, and kept the full eval free to re-run ($0) by recording model responses once and replaying them. One automated grader agreed with people too rarely (κ = 0.14), so its score was dropped rather than published.
  • Cut model cost 26–39% on 8 of 13 kinds of question with no drop in accuracy, by sending each question to the cheapest model that answers it well and falling back down a chain that ends in a person.
  • Found and fixed a ranking leak between companies inside the search engine, and pinned it with a test so it cannot come back.
  • FastAPI
  • Hybrid RAG
  • Qdrant
  • MCP
  • Langfuse
  • pytest
How I built it →

API

TypeScript

Forget the company filter on a query and you get nothing back, never someone else’s data.

  • Postgres itself blocks the rows, so a mistake in the code cannot leak them. Asking for another company’s ticket returns 404, not 403; saying “forbidden” would confirm the ticket exists.
  • NestJS
  • Prisma
  • PostgreSQL RLS
  • PgBouncer
  • Socket.IO
  • Redis
How I built it →

Front ends

TypeScript

Every screenshot above was taken from the running app by a script.

  • Four apps share one codebase: the agent dashboard, the customer portal, the analytics view, and a chat box that drops into any website under 60 KB without picking up that site’s styling.
  • Next.js 15
  • React 19
  • Tailwind v4
  • TanStack
  • Shadow DOM
  • Playwright

Nostro

Systems 2026

A ledger for many companies where money cannot go missing

A payments company has to show its books balance to the cent. Here the database rejects any entry that would break that, instead of a nightly check finding it later.

  • Java
  • Spring Boot
  • Kafka
  • gRPC
  • PostgreSQL
  • Testcontainers
Source Write-up
The full write-up
An entry that doesn’t balance, arrives twice, or touches another company’s money cannot be recorded. The database refuses it.

A double-entry ledger API for many tenants on one deployment, split into three Spring Boot services. Every guarantee in the README names the test that proves it, and CI ticks that table on every push.

  • 211 tests (127 against real Postgres, Kafka and Redis) prove the guarantees in a 3-minute CI run: 40 writers on one account, opposite-order writers, concurrent duplicates, and killing the relay’s leader mid-stream.
  • Load-tested a hot account: half the throughput and 2.3× the p99 on one account versus 64, with zero database errors either way. Every refusal was the balance floor, answered as a clean 422.
  • Failure cases are sealed types, so a failure nobody handled is a compile error, not a 500 in production.
Guarantee Every entry sums to zero, is recorded once, and stays inside its own tenant
Built with Transactional outbox · Kafka relay · gRPC read model · Postgres row-level security
Size 3 services · 211 tests · 16 decision records
Runs on docker compose up, which seeds two tenants and prints their keys

LinkPulse

Infrastructure 2026

A small app, run like a production platform, then broken on purpose on real AWS

An alert nobody has seen fire may stay silent in the outage it was written for. Every alert here was made to fire on purpose.

  • Go
  • Terraform
  • Kubernetes
  • AWS EKS
  • ArgoCD
  • Prometheus
  • Grafana
Source Write-up
The full write-up
Every alert has been seen to fire, because I made each failure happen.

A URL shortener with click analytics, where the product is small on purpose and the work is operating it: Terraform, Kubernetes, GitOps, monitoring, and chaos tests, on $0 plus one budgeted window on AWS.

  • Ran the whole stack for 12 hours on AWS EKS against real DynamoDB for $3.32, and found and fixed 13 defects no local run or CI could have shown.
  • Proved 13 alert rules with 5 chaos experiments (4 run in CI on every push). Two alerts turned out to be unable to ever fire; the write-up is a blameless postmortem.
  • With real DynamoDB throttling clicks, redirects stayed at 0 failures in 7,202 requests and a 7 ms p99, because clicks are written in the background and the redirect never waits for them.
Guarantee Redirects stay fast even when the database is throttling or down
Built with Terraform · Kubernetes on AWS EKS · ArgoCD · Prometheus, Grafana, Loki
Size 4 Terraform environments · 13 alerts · 5 chaos experiments · 9 CI jobs
Runs on An always-free AWS tier, and a local cluster with one command
Operator dashboard showing live counters, demo controls and the scenario buttons Operator dashboard showing live counters, demo controls and the scenario buttons
The operator dashboard
Event list showing dead-lettered events with every attempt recorded Event list showing dead-lettered events with every attempt recorded
Every attempt, including the failures

Webhook Delivery Engine

Systems 2026

Sends webhooks and never loses one, even when things break

Payment providers and SaaS tools tell other systems what happened through webhooks, and a lost one is a missed payment or order. This engine delivers each one or records why it couldn't.

  • Express
  • BullMQ
  • Redis
  • PostgreSQL
  • Docker
  • OpenAPI
Watch it work (20 s) Live demo API reference Source Write-up
The full write-up
Once the API says 202, the event is either delivered or recorded as failed, and every attempt in between is written down.

There is no fourth outcome where an event quietly disappears. That is the whole point, and the hardest part to keep true.

  • Postgres holds the truth; Redis is only a scheduler. Wiping all of Redis in the middle of a load test lost 0 of 50,000 events, sent none twice, and needed no restart: all 50,000 were delivered within 191 seconds.
  • The first time I ran that test it failed: 23,536 events sat undelivered until a restart, because the wipe also deleted the sweeper’s schedule. A watchdog now puts the schedule back, and the same test passes. Measure the claim, find it false, fix it, measure again.
  • Held 5,000 events a minute for 15 minutes at 47 ms p95 with a flat backlog, and turned 10,000 events sent twice into exactly 10,000 deliveries. Every delivery is signed, Stripe-style.
Guarantee Accepted once · delivered at least once · never silently lost
Built with Outbox pattern · dead-letter queue with replay · signed requests
Runs on $0 of hosting
MoviesWave home: a featured film and a grid of posters by category MoviesWave home: a featured film and a grid of posters by category
Popular, by genre
A film's page: poster, rating, overview, top cast and links A film's page: poster, rating, overview, top cast and links
One film, with its cast

MoviesWave

Side project 2023

A film browser that remembers what you already looked at

  • JavaScript
  • React
  • Redux
Live Source
The full write-up

Caching in the browser cut repeat calls to the film database by about half in a normal browsing session.

MetaMask ETH Bank

Side project 2023

Deposits and withdrawals on a smart contract, signed in with a wallet

  • Solidity
  • Ethereum
Source
The full write-up

A smart contract for deposits and withdrawals, using a crypto wallet to sign in. No server involved.

Toolkit things I have actually used

Languages

  • Python
  • TypeScript
  • Java
  • JavaScript
  • Node.js
  • SQ SQL
  • Go
  • PHP
  • C/C++

AI & LLM Engineering

  • LangGraph
  • MCP
  • RA RAG
  • LangChain
  • Qdrant
  • CH ChromaDB
  • EM Embeddings
  • EV Eval harnesses
  • GU Guardrails
  • LA Langfuse
  • PR Prompt engineering
  • MU Multi-model routing
  • VI Vision LLMs
  • VO Voice agents

Models

  • Anthropic Claude
  • OP OpenAI
  • Google Gemini
  • GR Groq
  • SA Sarvam AI
  • Ollama (Llama, Qwen, Gemma)

Security

  • CE CEH v11
  • OWASP LLM Top 10
  • Postgres RLS
  • PR Prompt-injection resistance
  • HM HMAC signing
  • JWT + RBAC
  • AP Applied cryptography
  • AD Adversarial testing

Backend

  • FastAPI
  • Spring Boot
  • NestJS
  • Express.js
  • GraphQL
  • gRPC
  • Kafka
  • OU Outbox pattern
  • ID Idempotency
  • BU BullMQ
  • Redis
  • Socket.IO
  • WS WSO2 SSO
  • OpenAPI

Data

  • PostgreSQL
  • MongoDB
  • MySQL
  • DynamoDB
  • Redis
  • Qdrant
  • Prisma
  • SC Schema design
  • QU Query optimization

Frontend

  • React
  • Next.js
  • Angular
  • Redux
  • TanStack Query
  • Tailwind CSS
  • RxJS

Infrastructure

  • Docker
  • Kubernetes
  • Terraform
  • ArgoCD
  • Helm
  • GitHub Actions
  • Jenkins
  • Prometheus
  • Grafana
  • Nginx
  • Linux
  • AW AWS (EKS, DynamoDB, S3, IAM)
  • Azure (App Service, Blob)

Testing & Quality

  • pytest
  • JUnit 5
  • TE Testcontainers
  • Jest
  • Vitest
  • Playwright
  • SU Supertest
  • MSW
  • k6
  • Gatling
  • Selenium
  • Postman

Integrations

  • Razorpay
  • IC ICICI Payments
  • DI Digilocker
  • SL Slack Bolt
  • Gmail API
  • Google Calendar
  • MI Microsoft Graph
  • Metabase

Blockchain & Mobile

  • Solidity
  • Ethereum
  • AV Avalanche
  • Polygon
  • ER ERC-20
  • Capacitor (iOS & Android)

Writing most recent

Nov 30 A message queue is a log that only grows and one number per consumer group. Every guarantee is a rule about when that number moves
low-level-designjavainterview-prep
28 min Nov 29 Stock is not a counter you decrement. It's three numbers, and every order moves units between them in one atomic step
low-level-designjavainterview-prep
26 min Nov 28 Splitwise is a ledger that must balance to the last paisa. Splitting the bill is the easy part
low-level-designjavainterview-prep
27 min

All writing

Credentials degree, certificates, awards

Education

  • B.E. Computer Science Engineering

    Specialization in Information Security

    Chandigarh University, Mohali·Aug 2020 to May 2024

    CGPA 8.02 / 10

  • Higher Secondary (12th), Science

    PCM + Computer Science

    Govt. Sr. Sec. School, Chotta Shimla·Mar 2019

    90% (HPBoSE)

Certifications & awards

  • Certified Ethical Hacker (CEH) v11

    EC-Council · 2023–2026

    Verify

  • React Basics & Advanced React

    Meta, via Coursera · 2023

    React Basics Advanced React

  • ETH Proof · ETH+AVAX Proof · Poly Proof

    Metacrafters · on-chain, Solana-minted

    ETH Proof ETH+AVAX Proof Poly Proof

  • Top 5 nationally · Intel oneAPI Hackathon

    Intel × IIT Roorkee.

  • District Rank 1 · Mathematics Olympiad

    National Science Congress.

Off-hours Chess · Technical writing · 10-finger typing · Infrastructure spelunking

Get in touch

rishabhsharma8912@gmail.com