Shiv Shankar Singh sh1v-max

👋 Hi, I'm Shiv

I'm a full-stack developer from Bengaluru, India who builds with React, Node.js, Express, MongoDB and PostgreSQL, and I've spent the last few months going deep on how AI backends actually work: embeddings, vector search, RAG, conversation memory, streaming and structured LLM output.

I like building things end to end and shipping them. Every project below is live, deployed on free-tier infrastructure, and has a README that explains the why behind the decisions, not just the features.

const shiv = {
  role: "Full-Stack Developer",
  basedIn: "Bengaluru, India 🇮🇳",
  education: "B.Tech CSE, Lovely Professional University (2024)",
  stack: ["TypeScript", "React", "Node.js", "Express", "PostgreSQL", "MongoDB"],
  ai: ["RAG", "Embeddings", "pgvector", "SSE streaming", "Structured output + Zod"],
  nowLearning: ["BullMQ + Redis", "Agents & tool calling", "Vitest + Supertest", "Docker"],
  consistency: "300+ day GitHub streak · 100+ days of DSA",
  openTo: ["Full-Stack", "AI Backend", "Frontend"],
};

🧠 Flagship: DocMind

DocMind: Chat With Your PDFs

Upload a PDF, chat with it, get quizzed on it. An AI backend built from scratch to understand how AI products work under the hood. No LangChain, no LLM SDK: just Express, Postgres and plain HTTP calls to Gemini, so every piece is visible.

TypeScript Express 5 PostgreSQL pgvector Drizzle ORM Gemini API Zod Server-Sent Events React 19

flowchart LR
    A[📄 PDF upload] --> B[Parse text]
    B --> C[~500-word chunks]
    C --> D[Gemini embeddings<br/>3072-dim]
    D --> E[(Postgres + pgvector)]
    Q[💬 Question] --> F[Embed question]
    F --> E
    E -->|top-3 by cosine distance| G[Prompt: chunks + last 8 turns]
    G --> H[Gemini]
    H -->|streamed over SSE| U[⚛️ React UI]

📄 Ingestion pipeline: PDF → text → chunks → Gemini embeddings (3072-dim) → pgvector, each chunk tagged with its document
🔍 RAG chat over one PDF or all of them, with the source passages (and which file they came from) shown under every answer
🧵 Conversation memory in Postgres: the last 8 turns go into every prompt, and a page refresh resumes the same chat
⚡ Live streaming over SSE, parsed by hand from Gemini's raw byte stream with an async generator
📝 Quiz generation with Gemini's structured-output mode, validated with Zod, retried once on bad output, options shuffled in code, graded in the browser
☁️ Deployed on free tier: API on Render, frontend on Vercel, Postgres on Neon

🔎 Engineering decisions worth knowing

Validate the model like any untrusted client. Every quiz goes through a Zod schema before it's used. The model was valid 17/17 times in testing, so it's insurance, and a gallery of deliberately broken replies proves what it catches.
Valid isn't the same as good. Across 50 generated questions the correct answer landed unevenly (one slot 36%, another 10%), so options are shuffled in code (Fisher–Yates) after validation instead of trusting the model's habits.
Retry once, then fail honestly. Two attempts max, then a clean 502. Never an endless loop.
Errors travel inside the stream. Once an SSE response starts, its status code is already sent, so failures go out as an in-band error event.
Readable errors. A failed vector query from Drizzle dumps all 3072 embedding numbers. A small helper prints just the message, the underlying cause, and the line in the codebase where it broke.
Routes → services → repositories. Only repositories touch SQL, and services never touch req/res, so the same pipelines can later run inside a background worker.

⏳ The API is on Render's free tier, so the first request after it's been idle can take up to a minute while it wakes up.

🚀 Featured Projects

🎬 Cinegraph

AI movie, TV & anime recommender

React 19 Redux Toolkit Firebase Gemini API Cloudflare Workers TMDB Tailwind v4

🔐 Gemini key kept server-side in a Cloudflare Worker proxy with a CORS allow-list and input caps
✅ Model returns strict JSON, and every title is cross-checked against TMDB, so no hallucinated picks reach the UI
🎯 Taste profile from your likes/dislikes (top genres, genres to avoid, favorite decade) fed into the prompt
💬 Multi-turn follow-ups ("more like the third one") and three personalized "For You" rows
🔥 Firebase Auth + Firestore with owner-only security rules

⚙️ TaskForge

Secured REST API + React task manager

Node.js Express 5 MongoDB JWT Zod Swagger React 19 GitHub Actions

🔐 JWT + bcrypt auth, and every query scoped to its owner, so users only ever see their own data
🛡️ Zod on bodies and query strings, Helmet, locked-down CORS, proxy-aware rate limiting (100 req / 15 min)
📄 Interactive Swagger / OpenAPI docs
🔍 Server-side filtering, sorting and pagination
🔁 GitHub Actions CI, auto-deploys to Render (API) and Vercel (frontend)

🥘 BharatDiet

Personalized nutrition for real Indian food

React 19 Vite Tailwind v4 React Router 7 Vitest

🍛 Meal planner matched to your calories, protein target, region, diet type and daily budget in ₹
🧮 Greedy meal allocator with a protein-boost pass and portion reconciliation ("2 roti → 2.5 roti")
📊 200+ Indian foods with real serving sizes, macros and cost, sortable by protein-per-rupee
🧪 Vitest suite: nutrition formulas, dataset validation, and every region × diet × budget × goal combination
🔓 Runs fully client-side: no account, no tracking

🍕 BiteSwift

Swiggy-style food delivery app

React 19 Redux Toolkit React Router Parcel Netlify Functions

🛒 Redux Toolkit cart with quantities, a bill breakdown (GST, fees, coupons) and a simulated checkout
🌐 Netlify Functions proxy for Swiggy's listing API, which sends no CORS headers
🧯 Menus blocked by bot protection fall back to mock data in the exact same shape, so the app never breaks
💡 Shimmer loading, lazy-loaded routes, custom hooks

🧩 More builds

Project	What it is	Stack	Links
🌐 Portfolio	6 themes (WCAG AA checked), live GitHub dashboard with a contribution calendar, project case studies, 33 UI experiments, contact form via a Netlify Function	React, Vite, Tailwind v4, Framer Motion	Live · Code
📚 BookVerse	Book discovery with debounced search by title, author or genre, trending lists and detailed book pages	React, Tailwind CSS, Open Library API	Live · Code
🗺️ Where Is Your Country	One of my early builds for exploring countries of the world	JavaScript	Live

🛠️ Tech Stack

Languages
Frontend
Backend & Data
DevOps & Tools
AI & Libraries
Learning now

🧭 How I Build

Understand it, then build the smallest real version. DocMind has no framework doing the interesting parts on purpose, so I know what RAG, streaming and structured output look like underneath.
Treat model output as untrusted input. Validate it with a schema, cross-check it against real data, and never let a hallucinated result reach the user.
Secrets stay on the server. API keys live behind a proxy or backend, never in the browser bundle.
Ship it, then document the why. Every project is deployed and has a README that explains the tradeoffs, including known limitations.
Free-tier first. Render, Vercel, Netlify, Neon, Firebase and Cloudflare Workers, no card required.

📊 GitHub Stats

🤝 Let's Connect

I'm open to Full-Stack and AI Backend roles. If you're building something with LLMs, APIs or real users, I'd love to talk.

Navigation Menu