Chetan Sharma

000%

chetan.os

Ch.00Prologue

Chetan Sharma

Blog

Resume

~/story/00_prologue

ChetanSharma

Designing the systems behind AI products.

A senior full-stack engineer in Bengaluru — three years building the load-bearing parts of AI products: auth, sandboxes, payments and RAG pipelines.

Begin the story

Available·—·12.97°N 77.59°E

Scroll to explore

~/story/01_origins# chapter one

Before the systems, a person.

I'm a full-stack engineer who likes the hard parts — secure auth, serverless execution, real-time data, and the kind of backend systems that stay fast under load.

The current chapter is 123ofAI, where I build the load-bearing parts of an AI-powered product — a serverless code-execution engine on Modal, a custom JWT auth system, a payments & coupon engine, and document-ingestion pipelines.

The chapters before that: Spring Boot microservices, real-time Kafka systems, and React / Next.js apps across healthcare and SaaS — always with an eye on performance, security, and clean architecture. I like owning features end-to-end, from data model to UI.

chetan@bengaluru — zsh

Based in: Bengaluru, India
Focus: Backend systems · AI infrastructure
Experience: 3+ years, senior scope
Currently: Full-Stack Dev @ 123ofAI

~/story/02_the-road# chapter two

Five roles. One throughline: own it end-to-end.

From building AI infrastructure in Bengaluru back to an internship in Noida — the road, latest mile first.

2025
Beat 01 — Present day
123ofAICurrent
Full-Stack DeveloperJun 2025 — Present · Bengaluru, India
Building the hard parts of an AI product: authentication, sandboxes, payments and pipelines.
4 systems shipped98.1% route coverageAuth0 → in-house
- Built a custom JWT auth system (HS256, 7-day tokens) with multi-device session management — a 4-session-per-user cap enforced via Azure Tables + Redis with idle-session reaping — migrating the app off Auth0.
- Engineered a serverless code-execution engine on Modal: a sandboxed Python runtime (512 MB / 1 CPU, 120 s cap, restricted builtins) shipping a preinstalled ML stack.
- Developed a Razorpay payments + coupon engine with O(1) synchronous redemption validation, multi-currency plans, and per-influencer conversion attribution.
- Built a 12-step DOCX → Markdown + LaTeX ingestion pipeline with deterministic math-block restoration, plus a Playwright e2e suite and 98.1% backend route coverage.
Next.jsTypeScriptNode.jsModalPythonAzureRedisPlaywright
2024
Beat 02 — The solo flight
Healthcare Appointment Portal
Full-Stack Engineer (Freelance)Aug 2024 — Jan 2025 · Remote · Canada
A clinic platform built end-to-end, alone — from data model to SMS reminders for 1,500+ monthly users.
1,500+ monthly users30% no-shows ↓25% faster queries
- Built a clinic scheduling platform with Spring Boot and React, serving 1,500+ monthly users.
- Cut no-shows 30% with automated Twilio SMS reminders; secured access with JWT/RBAC.
- Improved database speed 25% via MySQL indexing and Redis caching; integrated Apache Kafka and deployed on AWS EC2/S3.
Spring BootReactMySQLRedisKafkaTwilioAWS
2024
Beat 03 — Going distributed
Globalogy
Full-Stack Development EngineerFeb 2024 — Jun 2024 · Remote · Canada
First taste of scale: microservices, Redis, Kafka — and the discipline distributed systems demand.
90% latency ↓40% infra cost ↓50% deploy time ↓
- Developed end-to-end Spring Boot microservices with JWT auth, contributing to a 40% infrastructure cost reduction.
- Cut latency 90% with Redis caching and Kafka-based inter-service communication.
- Containerized the application ecosystem with Docker, reducing deployment time 50%.
Spring BootRedisKafkaDockerMicroservices
2022
Beat 04 — Learning to ship
Wathare Infotech Solutions
Full-Stack Development EngineerJul 2022 — Jan 2024 · Pune, India
Eighteen months of React, Node and MySQL — learning what production actually demands of code.
- Built and maintained web applications with React, Node.js, and Express, improving engagement and functionality.
- Designed RESTful APIs and optimized MySQL queries, indexes, and joins to reduce response times.
ReactNode.jsExpressMySQLREST APIs
2022
Beat 05 — Where it started
Openwork Park
Junior Developer InternJan 2022 — Jun 2022 · Noida, India
Six months of AngularJS and front-end performance — enough to learn that shipping beats theory.
- Revamped website designs in AngularJS and optimized front-end performance, reducing page load times.
AngularJSJavaScript

Interlude ·The scoreboard so far

latency cut with Redis + Kafka

backend route test coverage

monthly users served, built solo

infrastructure cost reduction

~/story/03_the-builds# chapter three

The systems worth telling.

Four production systems designed and shipped at 123ofAI — the problem, the architecture, and the trade-offs behind each.

Build 01 / 04 — 123ofAI · Platform Infrastructure

Serverless Code-Execution Engine

A secure, low-latency engine that runs untrusted user code in ephemeral sandboxes on Modal — no standing servers, scales to zero.

512MB sandbox cap

120s hard timeout

0 standing servers

01
Run untrusted code, safely
Users submit arbitrary Python to execute remotely. The hard part is isolation, latency, and resource limits — without paying for standing infrastructure.
02
Authenticated dispatch
Submissions hit a Next.js API route that authenticates the user, enforces quotas, and dispatches the job to Modal.
03
Ephemeral sandbox
Modal spins up an isolated sandbox capped at 512 MB / 1 CPU with a restricted-builtins namespace — created on demand, scaled to zero when idle.
04
Execute with an ML stack
Code runs against a preinstalled stack (PyTorch, TensorFlow, Transformers, LangChain, FAISS) under a 120-second hard timeout.
05
Return & tear down
stdout streams back to the client and the sandbox is destroyed — clean, stateless, and cost-efficient by design.

serverless_code_execution_engine.arch compute data

Client

browser / editor

API Route

Next.js · auth + quota

Modal Dispatch

serverless

Sandbox

512MB · Python 3.10

Result

stdout · torn down

Build 02 / 04 — 123ofAI · Identity

Multi-Device JWT Auth

A custom authentication system that replaced Auth0 — JWT sessions with a hard device cap, backed by Azure Tables and Redis.

4 device cap

HS256 signed tokens

Redis session cache

01
Off Auth0, in-house
Migrated authentication to a custom JWT layer for full control over sessions and cost.
02
Issue tokens
Signed HS256 tokens with a 7-day lifetime are issued on login.
03
Track sessions
Active sessions live in Azure Tables, cached in Redis for fast lookups.
04
Cap to 4 devices
A hard 4-session-per-user cap is enforced on every login.
05
Reap idle sessions
Idle sessions are auto-reaped after 3 days, keeping the store clean.

multi_device_jwt_auth.arch compute data

Client

Auth API

credentials

JWT · HS256

7-day tokens

Session Store

Azure Tables + Redis

Device Cap

4 max · idle reaper

Build 03 / 04 — 123ofAI · Data

DOCX → Markdown + LaTeX Pipeline

A 12-step ingestion pipeline that converts Word documents to clean Markdown while preserving math with deterministic LaTeX restoration.

12 pipeline steps

100% math fidelity

WebP image output

01
DOCX in, Markdown out
Convert author documents to clean Markdown without losing equations.
02
Parse to HTML
Mammoth converts the DOCX into structured HTML.
03
Protect the math
LaTeX blocks are extracted to placeholders before conversion — so nothing mangles them.
04
HTML → Markdown
Turndown produces deterministic Markdown from the HTML.
05
Restore & optimize
Math is swapped back in and images are compressed to WebP.

docx_markdown_latex_pipeline.arch compute data

Upload

DOCX

Mammoth

DOCX → HTML

Math Extract

→ placeholders

Turndown

HTML → Markdown

Restore + WebP

math back · images

Build 04 / 04 — 123ofAI · Payments

Payments + Coupon Engine

A Razorpay-backed checkout with O(1) coupon validation, multi-currency plans, and per-influencer conversion attribution.

O(1) coupon lookup

₹/$ multi-currency

per-ref attribution

01
Real-time discounts
Coupons are validated instantly at checkout, with no external call in the hot path.
02
O(1) validation
A single Azure Table read resolves a coupon — constant time, no scans.
03
Four checks
Active, expiry, max-uses, and plan-tier are validated before applying.
04
Charge via Razorpay
Multi-currency plans (₹ / $, monthly & annual) are processed through Razorpay.
05
Attribute conversions
Each conversion writes a per-influencer attribution row with commission.

payments_coupon_engine.arch compute data

Client

checkout

Checkout API

validate + price

Coupon Store

Azure Table · O(1)

Razorpay

multi-currency

Attribution

per-influencer rows

Keep scrolling — the shelf slides

Build 01 / 04 — 123ofAI · Platform Infrastructure

Serverless Code-Execution Engine

A secure, low-latency engine that runs untrusted user code in ephemeral sandboxes on Modal — no standing servers, scales to zero.

512MB sandbox cap

120s hard timeout

0 standing servers

01
Run untrusted code, safely
Users submit arbitrary Python to execute remotely. The hard part is isolation, latency, and resource limits — without paying for standing infrastructure.
02
Authenticated dispatch
Submissions hit a Next.js API route that authenticates the user, enforces quotas, and dispatches the job to Modal.
03
Ephemeral sandbox
Modal spins up an isolated sandbox capped at 512 MB / 1 CPU with a restricted-builtins namespace — created on demand, scaled to zero when idle.
04
Execute with an ML stack
Code runs against a preinstalled stack (PyTorch, TensorFlow, Transformers, LangChain, FAISS) under a 120-second hard timeout.
05
Return & tear down
stdout streams back to the client and the sandbox is destroyed — clean, stateless, and cost-efficient by design.

serverless_code_execution_engine.arch compute data

Client

browser / editor

API Route

Next.js · auth + quota

Modal Dispatch

serverless

Sandbox

512MB · Python 3.10

Result

stdout · torn down

Build 02 / 04 — 123ofAI · Identity

Multi-Device JWT Auth

A custom authentication system that replaced Auth0 — JWT sessions with a hard device cap, backed by Azure Tables and Redis.

4 device cap

HS256 signed tokens

Redis session cache

01
Off Auth0, in-house
Migrated authentication to a custom JWT layer for full control over sessions and cost.
02
Issue tokens
Signed HS256 tokens with a 7-day lifetime are issued on login.
03
Track sessions
Active sessions live in Azure Tables, cached in Redis for fast lookups.
04
Cap to 4 devices
A hard 4-session-per-user cap is enforced on every login.
05
Reap idle sessions
Idle sessions are auto-reaped after 3 days, keeping the store clean.

multi_device_jwt_auth.arch compute data

Client

Auth API

credentials

JWT · HS256

7-day tokens

Session Store

Azure Tables + Redis

Device Cap

4 max · idle reaper

Build 03 / 04 — 123ofAI · Data

DOCX → Markdown + LaTeX Pipeline

A 12-step ingestion pipeline that converts Word documents to clean Markdown while preserving math with deterministic LaTeX restoration.

12 pipeline steps

100% math fidelity

WebP image output

01
DOCX in, Markdown out
Convert author documents to clean Markdown without losing equations.
02
Parse to HTML
Mammoth converts the DOCX into structured HTML.
03
Protect the math
LaTeX blocks are extracted to placeholders before conversion — so nothing mangles them.
04
HTML → Markdown
Turndown produces deterministic Markdown from the HTML.
05
Restore & optimize
Math is swapped back in and images are compressed to WebP.

docx_markdown_latex_pipeline.arch compute data

Upload

DOCX

Mammoth

DOCX → HTML

Math Extract

→ placeholders

Turndown

HTML → Markdown

Restore + WebP

math back · images

Build 04 / 04 — 123ofAI · Payments

Payments + Coupon Engine

A Razorpay-backed checkout with O(1) coupon validation, multi-currency plans, and per-influencer conversion attribution.

O(1) coupon lookup

₹/$ multi-currency

per-ref attribution

01
Real-time discounts
Coupons are validated instantly at checkout, with no external call in the hot path.
02
O(1) validation
A single Azure Table read resolves a coupon — constant time, no scans.
03
Four checks
Active, expiry, max-uses, and plan-tier are validated before applying.
04
Charge via Razorpay
Multi-currency plans (₹ / $, monthly & annual) are processed through Razorpay.
05
Attribute conversions
Each conversion writes a per-influencer attribution row with commission.

payments_coupon_engine.arch compute data

Client

checkout

Checkout API

validate + price

Coupon Store

Azure Table · O(1)

Razorpay

multi-currency

Attribution

per-influencer rows

~/story/04_the-craft# chapter four

Tools collected along the way.

Every chapter left something in the toolbox. Backend-leaning full-stack with real AI-infrastructure depth — distributed services, shipped to the cloud.

AI & LLM

RAG systems, agents, and the infrastructure they run on — the current frontier of the work.

LangChainLlamaIndexOpenAIClaudeHugging FacePyTorchRAGVector DBsAI Agents

Backend

The home turf: services that stay correct, fast, and secure under load.

Node.jsExpressSpring BootREST APIsMicroservicesJWT / OAuth

AI Infra & Data

ModalServerlessData PipelinesApache Kafka

Languages

TypeScriptJavaScriptJavaPythonSQLC++

Frontend

ReactNext.jsReduxTailwind CSSAngular

Cloud & DevOps

AzureAWSDockerCI/CDGit

Databases

MySQLMongoDBRedis

Testing

PlaywrightJestJUnitPostman

~/story/05_projects# chapter five

Built after hours.

Products, platforms and experiments shipped end-to-end outside the day job — each with the system design behind it.

Project 01 — AI-powered portfolio — this site

engineerchetansharma.com + RAG Assistant

A grounded RAG chatbot answers recruiter questions from my real career data — Gemini + Supabase pgvector, with citations on every claim.
Prompt-injection hardening, two-tier per-IP rate limiting (burst + daily), and deterministic refusal that skips the model entirely when retrieval finds nothing.
WebGL particle universe (Three.js): a neural galaxy hero with a big-bang intro that morphs through five formations as the story scrolls.
Self-publishing platform: section-based blog CMS and resume manager behind a secret-keyed admin, stored on Supabase, live via ISR — no redeploys.

Next.jsTypeScriptThree.jsGSAPGeminiSupabasepgvectorUpstash

Live view GitHub

portfolio-rag-assistant.applive

engineerchetansharma.com + RAG Assistant

portfolio-rag-assistant.arch — system design diagram coming soon

Project 02 — RAG built from scratch

Paperbase — a queryable home for your documents

Semantic search over personal documents (lease, insurance, passport, warranties) with every answer cited to its source file.
No LangChain, no vector DB, no LLM API — the whole retrieval pipeline is plain Python I can walk through line by line.
Local MiniLM embeddings with a pure-Python fallback; L2-normalized vectors make cosine search a plain dot product.
Local-first by design: the hosted demo runs fictional samples — the same code runs fully offline on your real documents.

Try it: “When does my passport expire?”

Pythonsentence-transformersStreamlitRAGVector Searchpypdf

Live view GitHub Read the build log

paperbase.applive