AK · Aman Keshri, home
  • HomeStart
  • WorkCase studies & projects
  • AboutStory, timeline
  • BlogNotes & articles
  • ContactStart a project
  • ServicesWhat you can hire me for
  • Tech stackTools I ship with
  • Brief builderPick what you need
  • Hardware Store ERPCase study I
  • Procure-to-Pay platformCase study II
  • VartalaapCase study III
  • Shiv Cement StoreBusiness platform · applied AI
  • AI-Powered Alexa AssistantVoice · applied AI
  • MedScan AIVision · applied AI
  • Mythos AI StudioCreative tools
  • Foxpop.inCommerce
  • URL ShortenerWeb utility
  • Prompt EnhancerLanguage · applied AI
  • Serverless Notification SystemCloud infrastructure
  • Resume (PDF)Download
  • RSS feedSubscribe

↑ ↓ to move · Enter to open · Esc to close

Hire me

WORK ✺ CASE STUDIES & PROJECTS

Systems that hold, and the things around them.

Three systems I built, each with a working model you can poke, break and reset. Below them, everything else: business platforms, applied AI, cloud infrastructure and small sharp tools.

CASE Nº I ✺ RETAIL & WHOLESALE BILLING

Hardware Store ERP: Everything goes through. Or nothing does.

Hardware Store ERPSole engineerMar 2026 — present● In production

A sale is never just a PDF. It moves stock, changes what a customer owes and lands in the books. This ERP keeps those three records in agreement — even when something fails halfway.

Visit the platform

Try it: sell a bag, then pull the plug halfway. Or race two counters for the last bag.

Instrument · one sale, three records

Counter Aidle
Counter Bidle
Invoices
  • No sales yet
Stock · OPC-53 cementunlocked

3 bags

Customer ledger

₹0 owed

  1. Sell a bag, or race two counters for the last one.
  1. One transaction, six steps

    Invoice creation runs as a single PostgreSQL transaction with FOR UPDATE row locks. Invoice, stock and ledger commit together or not at all — two counters can sell the last bag of cement without both succeeding.

  2. Slow work stays out of the sale

    BullMQ workers on Redis render invoice PDFs with Puppeteer and store them privately in S3, so the counter never waits on a document. pg_trgm search finds products from partial or misspelled names.

  3. Run where the shop runs

    Billing, inventory, pricing, unit conversion, customer ledgers and financial reports — deployed on Ubuntu with AWS RDS PostgreSQL and CloudFront.

CASE Nº II ✺ ENTERPRISE PROCUREMENT WORKFLOW

Procure-to-Pay platform: Check the role. Check the state. Then move.

Procure-to-Pay platformFull-stack engineer, CodeAptoJun 2026 — present● Active development

From a purchase request to a released purchase order, money only moves through gates. Each transition asks two questions before it happens: who is asking, and is this the right moment?

Try it: release the PO as the requester. The server refuses. Then walk it through the gates.

Instrument · guarded transitions

  1. Draft
  2. Submitted
  3. Approved
  4. Awarded
  5. Released

PR-2041 · 40 bags OPC-53 · owned by Tenant A

Signed in as
Tenant
  1. Pick a role, then try to move the request. Try releasing the PO as the requester.

An illustration of the pattern. Roles and rules are simplified; no CodeApto data.

  1. The whole workflow

    Front-end features and backend flows across requisitions, approvals, sourcing, supplier onboarding, award decisions and purchase-order release.

  2. Invalid transitions can't happen

    Workflow paths reject invalid status changes, duplicate actions and unauthorised PO release — enforced on the server, with tests around the main paths.

  3. Boundaries are part of the feature

    Role-based access and tenant-specific data handling are designed in, not bolted on. This account uses only public, resume-level information about the product.

CASE Nº III ✺ SERVERLESS REAL-TIME MESSAGING

Vartalaap: One recipient goes quiet. Everyone else still hears.

VartalaapIndependent project● Built and deployed

A real-time chat and calling platform on AWS-managed services. A message to a group is many deliveries, and one closed connection must not cancel the rest.

Try it: knock B offline and send. Then switch to Promise.all and send again.

Instrument · independent deliveries

yousendMessageLambdafan-outPromise.allSettledAconnectedBofflineCconnected
Connected
  1. B is offline. Send a message, then switch to Promise.all and send again.
  1. Connections with identity

    Cognito-authenticated WebSocket connections carry one-to-one and group messages, with audio and video calls through the Chime SDK and push notifications.

  2. Shaped by access patterns

    A DynamoDB single-table design follows how conversations are actually read. Fan-out uses Promise.allSettled, so a failed recipient is recorded rather than thrown.

  3. Carried into operations

    The same habits from Technosys — EC2, Route 53, IAM, security groups, Docker and GitHub Actions — make releases repeatable and rollback boring.

MORE PROJECTS ✺ 8 SHIPPED

Every project, frame by frame.

Search by tool, or open a card for how it’s built.

8 of 8 projects

  1. Shiv Cement Store website home page with product categories

    BUSINESS PLATFORM · APPLIED AI · 2025 — 2026

    Shiv Cement Store

    Customer website, staff and admin mobile app, and the REST API between them for a cement retailer in Patna.

    How it’s built

    A multilingual RAG assistant answers in Hindi, English and Hinglish. Staff check in with GPS and a rotating QR code; uploads go to S3 behind role-based auth.

    Next.jsExpressMongoDBExpoOpenAIAWS S3
    Architecture diagram of the Shiv Cement platform: web, mobile, API, database and AI assistantOpen the architecture diagram ↗Link to this project
    Open Shiv Cement Store
  2. Alexa assistant project artwork

    VOICE · APPLIED AI · DEC 2025

    AI-Powered Alexa Assistant

    A serverless Alexa skill that holds a contextual conversation through OpenAI.

    How it’s built

    AWS Lambda and Node.js, session memory in DynamoDB, logs in CloudWatch and secrets in Systems Manager Parameter Store.

    AWS LambdaNode.jsOpenAIDynamoDB
    Link to this project
    Open AI-Powered Alexa Assistant
  3. MedScan AI interface identifying a medicine from a photo

    VISION · APPLIED AI · EARLIER WORK

    MedScan AI

    Photograph a medicine strip and get a readable summary, using Gemini Vision. Web, mobile and an Android APK.

    How it’s built

    An information tool. Its output is not a substitute for professional medical advice.

    Gemini VisionAWSNext.jsReact Native
    Link to this project
    Open MedScan AI
  4. Mythos AI Studio interface turning a prompt into a narrated video

    CREATIVE TOOLS · EARLIER WORK

    Mythos AI Studio

    Turns a prompt into a narrated video through a small team of cooperating agents.

    How it’s built

    Image generation and video composition behind a Streamlit interface.

    PythonStreamlitStable DiffusionMoviePy
    Link to this project
    Open Mythos AI Studio
  5. Foxpop makhana store promotional banner

    COMMERCE · EARLIER WORK

    Foxpop.in

    An online store for makhana: catalogue, payments and a deployment pipeline.

    How it’s built

    MERN on AWS EC2 behind Nginx, with Razorpay for payments.

    MongoDBExpressReactNode.jsAWS EC2NginxRazorpay
    Link to this project
    Open Foxpop.in
  6. URL shortener input form

    WEB UTILITY · FEB 2026

    URL Shortener

    Long link in, short link out. It also powers the six-character links on this domain.

    How it’s built

    Next.js and TypeScript on Vercel, mounted under amankeshri.com through rewrites.

    Next.jsTypeScriptVercel
    Link to this project
    Open URL Shortener
  7. Prompt Enhancer interface

    LANGUAGE · APPLIED AI · EARLIER WORK

    Prompt Enhancer

    Rewrites a rough language-model prompt into clearer instructions.

    How it’s built

    A focused interface around the OpenAI API, built with Next.js.

    Next.jsOpenAIVercel
    Link to this project
    Open Prompt Enhancer
  8. Architecture of the serverless notification system on AWS

    CLOUD INFRASTRUCTURE · JAN 2026

    Serverless Notification System

    Event-driven email and SMS notifications, with Step Functions coordinating state and delivery.

    How it’s built

    Lambda, API Gateway, SNS and SES — cost-conscious and with nothing to keep running between events.

    AWS LambdaStep FunctionsSNSSESPython
    Link to this project
    Architecture project, no public demo

CONTACT ✺ ONE EMAIL AWAY

Bring me the problem.

Send a short brief. I reply with a plan, a timeline and a price.