Design Engineer · Applied AI
I build AI products, then design the brand story that sells them.
Portfolio · LinkedIn · Email · Resume
Guna, MP · open to relocate · B.Tech CSE 2026
I start from the client's need, ship the full-stack AI tool or app, and then turn it into a brand story people can read in 30 seconds: a case study, a launch poster, a banner and a narrated explainer video. One person, from problem to launch.
| Step | What happens | |
|---|---|---|
| 01 | Listen | Write the client's problem in one line, and who loses money or time because of it. |
| 02 | Build | Ship the AI tool, web app or Android app: UI, backend, AI layer, payments, deploy. |
| 03 | Guard | AI drafts, people decide. One gateway, spend caps, evals and load tests. |
| 04 | Brand | Name, colours, poster, banner and a QR code that opens the live demo. |
| 05 | Tell | Case study and narrated explainer video, so a buyer gets it in 30 seconds. |
- 500/500: Stack: requests served with the top two AI providers down
- 100%: CloseBy: over-cap AI calls blocked before any model call
- 27: CloseBy: eval cases, Hindi aliases to gibberish
- 7: ScanCart.ai: parallel checks behind one Trust Score
- 600+: teams at Hacksagon 2025, where I was a finalist
01 · CloseBy
Your street's shops, online. AI web app · marketplace · Solo: design, frontend, backend, AI layer · 2026
Neighbourhood shop owners have no time to type a catalogue, and buyers can't see what a shop near them has in stock.
- Owners import stock from a shelf photo or a voice list. The AI drafts; nothing is saved until the owner approves.
- Every AI call passes one gateway: auth, rate limit, kill switch, daily spend cap, tool loop, cost logging.
- Razorpay holds stock until the server confirms payment; rejected paid orders refund automatically.
100% over-cap AI calls blocked under load · 27 eval cases (Hindi, typos, gibberish) · 0 oversells: 50 buyers, last 5 units
Built with: Next.js · TypeScript · Supabase · Claude API · Razorpay · JMeter
Case study · Live demo · GitHub
02 · Stack
Stack makes you remember what you read. Web + Android app · AI reading · Solo: product, design, Android app, sync, AI layer · 2026
People read articles and PDFs all day and forget almost all of it a week later. Read-later apps save links but never bring them back.
- One Next.js codebase ships the website and a native Android app (Capacitor). Works offline in IndexedDB, no account needed.
- Every morning, 3 old highlights come back on a spaced schedule.
- Stack AI runs in one Supabase Edge Function and falls back across five model providers.
500/500 requests served with top 2 providers down · 1.6 s p95 latency in that failover test · 5 LLM providers in the fallback chain
Built with: Next.js 16 · TypeScript · Capacitor · IndexedDB · Supabase Edge Functions
Case study · Live app · GitHub
03 · ScanCart.ai
One Trust Score for any listing. AI Chrome extension + Android app · trust · Solo: extension, Android app, backend, agents · 2026
Shoppers on Indian marketplaces can't easily tell a fake product or a paid review from a real one.
- A Chrome extension sends the listing to a backend that runs 6 specialist AI agents in parallel, then a compliance agent.
- All 7 checks combine into one weighted Trust Score with the reasons behind it.
- A check that fails is left out and shown as "X of 7 checks", never guessed.
- Trust-first scoring: a real MacBook listing that was wrongly flagged (63, caution) now scores 86, safe; a fake ₹24,999 MacBook scores 40, do not buy.
- An Android app brings the same panel to phones: share a link to ScanCart, tap a floating bubble, or turn on the automatic check inside shopping apps.
7 checks behind one verdict · 1× latency = slowest check, not the sum · Top 35 finalists, Pitch-a-Thon 2026 Jaipur
Built with: Node.js · Express · Claude API · Chrome MV3 · React · Android
Case study · Website · GitHub
04 · Streetwise
One QR on every shop. One AI agent behind it. Multi-agent AI · local buying · Solo (Team SOLOX): agents, messages, maps, payments · 2026
To buy locally, people walk or call shop after shop to ask what is in stock and what it costs. Hidden fees show up only at the counter.
- Four AI agents: Shopper, Store Locator, one Merchant per shop, and Trust.
- The LLM only reads the shopping list. Prices, totals and fee checks are plain code, so it can't invent a number.
- Search widens 500 m → 1 km → 2 km until 3 open shops are found on OpenStreetMap.
4 AI agents talking in typed messages · 3 options: cheapest, nearest, one shop · 0.002 FET agent charge, verified on testnet
Built with: Python · Fetch.ai uAgents · ASI1 LLM · OpenStreetMap · Leaflet
Built for Fetch-A-Thon 3.0. Tested end to end locally; prices and orders are demo data.
05 · Orderly
Run your kirana shop on WhatsApp. Chat to cash, no app. WhatsApp bot + owner app · Solo: owner app, WhatsApp bot · 2026
Kirana customers already order over WhatsApp, but owners reply by hand and lose track of orders.
- An always-on WhatsApp bot answers price and stock questions from the live catalogue.
- Owner replies typed in the app are relayed back to WhatsApp.
- A crash supervisor and a 45-second heartbeat show the bot's health in the owner app.
45 s bot heartbeat, live in the owner app · 0 apps the buyer has to install · ₹ paise money stored as integers, never floats
Built with: React · Supabase · Clerk · Node.js · BuilderBot
In build: bot and owner app work; the payment-link flow is planned.
| Project | Test | Load | Result |
|---|---|---|---|
| Stack | Healthy | 100 users × 5 summaries | 500/500 OK, p95 817 ms |
| Failover | Sarvam + Gemini return 503 | 500/500 served by Groq, p95 1.6 s | |
| Total outage | All engines down, 100 requests | 100/100 got a clean error message, no crash | |
| Daily-limit race | 80 simultaneous requests, one person, limit 50 | Counter stopped at exactly 50; the rest got 429 | |
| Bad sign-in | 50 forged tokens | 50/50 rejected with 401 | |
| CloseBy | Last-stock race | 50 buyers at the same instant, 5 units left (×4 runs) | Exactly 5 orders, 45 clean rejections, stock 0: 0 oversells in 4/4 runs |
| Checkout load | 100 buyers × 8 COD orders | 800/800 OK, 0 stock mismatches, ~160 orders/s, p95 697 ms | |
| Nearby search | 50 users, 30 s | 10,529 requests, 0 errors, ~350 req/s, p95 153 ms | |
| Nearby search | 200 users, 60 s | 21,690 requests, 0 errors, ~360 req/s, p95 620 ms | |
| AI rate limit | 1 user, 30 rapid calls | 15 allowed, 15 rejected with 429 | |
| AI spend cap | User over the daily cap, 20 calls | 20/20 blocked, 0 model calls made | |
| AI gateway load | 50 users × 5 calls | 250/250 OK | |
| Forged Razorpay webhooks | 100 bad signatures | 100/100 rejected with 401 | |
| ScanCart.ai | Baseline | 1 user × 20 scans | Avg 4.8 s vs 13.7 s total model time per scan: 2.8× faster |
| Concurrency | 50 users × 4 scans | 200/200 OK, avg 4.76 s (same as 1 user), p95 5.8 s | |
| Hung agent | Review agent never answers, 20 scans | 20/20 still scored in ~12 s (10 s timeout, then fallback) | |
| Validation | 50 bad payloads | 50/50 rejected with 400 | |
| Rate limit | 120 calls from one client | 100 allowed, then 429 |
Stack: JMeter 5.6.3 against the real Deno edge function and the real ai_take() SQL; AI engines replaced by mock servers with realistic delays. CloseBy: JMeter 5.6.3 against a production build (1 Node process, 5-connection DB pool) with local Postgres, mock Claude and stubbed Clerk users. ScanCart.ai: JMeter 5.6.3 against the Express backend with all seven agents; Claude replaced by a mock server answering in 1–3 s per call.
| Tool | What it is | Status |
|---|---|---|
| Wakeface | A flip clock that is also my lock screen and phone app. Android app · clock + lock screen | v2.0, in daily use on my phone. Not published. |
| OTG Switch | One tap to turn on USB OTG, from the home screen. Android app · one-tap OTG | v3.0, in daily use on my phone. Not published. |
| GitHub Widgets | My GitHub contribution calendar and repos on my home screen. Android home-screen widgets | In daily use on my phone. Not published. |
| Mac Bar | A macOS-style top bar and dock for Ubuntu. GNOME Shell extension | Open source, part of Ubuntu Mac Kit. |
| Mac Notch | A live glass notch and a Mac-type lock screen for Ubuntu. GNOME Shell extension | In daily use on my desktop. Not published. |
| Mac Snap | Mac-style window snapping for Ubuntu. GNOME Shell extension | Built for my desktop. Not published. |
A complete macOS-style desktop on Ubuntu, in one installer. Ubuntu 24.04+ · GNOME 46 · v1.1.0 · open source
- WhiteSur-dark icons with regenerated blue macOS folders and white page file icons
- macOS cursor, dark GTK styling, window buttons on the left
- Mac Bar: blurred top bar, Ubuntu logo menu, dock magnify + bounce
- Sticky notes, optional Desktop Stack pile and GitHub widget
- Backup on install; the uninstaller restores your old settings
tar -xzf ubuntu-mac-kit.tar.gz && cd ubuntu-mac-kit && ./install.shDownload ubuntu-mac-kit.tar.gz from the latest release first. No sudo; everything goes into your home folder.
Glass GitHub widgets for the Ubuntu desktop. Ubuntu 24.04 · v1.0.1 · MIT licence
- GitHub Repo Analyser: pick a repo, see commits per day and replay its latest merged pull request
- GitHub Contribution Calendar: your contribution grid on the desktop
- Glass cards that sit on the desktop and toggle on and off
curl -fsSL https://github.com/chhari07/ubuntu-widgets/releases/latest/download/install.sh | bashOne command. Installs both widgets into your home folder.
- Guardrails first. Every model call passes one gateway with auth, rate limits, a kill switch and a daily spend cap.
- AI drafts, people decide. The model proposes stock imports, replies and restock ideas. Nothing is saved until the owner approves it.
- Checked output. Responses are validated against a Zod schema. Malformed output fails closed: the user sees an error and nothing is saved.
- Measured, not guessed. Eval cases cover misspellings, Hindi aliases and gibberish; JMeter shows 50 buyers racing for the last 5 units still make exactly 5 orders.
AI drafts, humans confirm. · State machines over vibes. · Money as integer paise, never floats. · A failed check is left out, never guessed.
| Roles | Design Engineer · Creative Technologist · AI Product Engineer · Forward Deployed Engineer |
| Type | Full-time |
| Level | Entry · B.Tech CSE 2026 |
| Location | Open to relocate |
| Start | Immediately |
- Design Engineer: I design the screens and build them, from Figma to a deployed Next.js app.
- Creative Technologist: I use AI and code to make campaigns, posters, interactive pages and explainer films.
- AI Product Engineer: I take a client's need and ship the whole AI product, with guardrails, evals and a clear story.
- Forward Deployed Engineer: I sit with the client, build the custom AI solution, and hand it over as a case study.
- Design: Figma, Brand & posters, Explainer video, CSS motion, Design systems
- Frontend: Next.js, React, TypeScript, Tailwind CSS
- Backend: Node.js, Express, FastAPI, Python, Supabase Postgres, Clerk, Razorpay
- Applied AI: Claude API, OpenAI API, Agentic workflows, Fetch.ai uAgents, Tool use, Structured outputs (Zod), Guardrails, Human-in-the-loop, LLM evals
- Infra & testing: Vercel, GitHub Actions, Linux, Apache JMeter, Vitest
Freelance Full-Stack & AI Developer, Independent · Guna, MP (Remote) Feb 2023 – Present
- Rent It, vehicle rental platform Aug 2025: React and Tailwind app: browse vehicles, book and pay through Razorpay, Firebase sign-in, protected routes, map section and booking history.
- Gurudev Public School, Guna Oct 2025: The school's 15-page website in React, Vite and Tailwind: academics, facilities, gallery, news, holiday homework and the public disclosure page.
Web Development Intern, Zidio Development · Remote Aug – Sep 2024
B.Tech, Computer Science & Engineering, Jaypee University of Engineering & Technology, Guna 2022 – 2026
- Finalist · 600+ teams, Hacksagon 2025 · ABV-IIITM Gwalior (certificate)
- Selected pitch · Top 35, Rajasthan AI Builders Pitch-a-Thon 2026 · Jaipur · ScanCart.ai (certificate)
Text generated from the portfolio's content files by node scripts/build_readme.mjs, so it matches amankumarchhari.vercel.app.


