I'm a Computer Engineering student and full-stack developer. I recently completed a Software Developer internship at FirstBookIt — a live sports-venue booking product on the web, Google Play and the App Store — where I owned modules end to end: the booking engine, split payments, the owner/manager/admin portals, auth and backend performance.

I care about shipping things that actually work — clean architecture, production deployments, and code that solves a real problem, not just a demo.

• 🔭 Right now: preparing for the Odoo × LDCE Hackathon '26 24-hour finale (3–4 Oct), and building Vajra — an offline, autonomous find → patch → prove security system for Terrier Cyber Quest 2026
• 🤖 Exploring: AI-powered products using LLM APIs (Groq, Anthropic/Claude)
• 🏗️ Shipped: 25+ production features in a completed internship · 7 live client websites · 19+ builds across 16 industries
• 🤝 Building with teams: on-site finals at Odoo India HQ (Odoo Hackathon 2026), ISRO's Bharatiya Antariksh Hackathon, Smart India Hackathon 2026, the Gujarat Police Innovation Challenge and SEMICON India — shared repos, PR-reviewed workflows, commits into teammates' repositories
• 🌱 Deepening: Data Structures & Algorithms (Java) and system-design fundamentals
• 📫 Reach me: het@buildbyhet.me · buildbyhet.me

• 🔧 Full-Stack Development — MERN stack (MongoDB, Express, React, Node.js), Next.js, REST APIs, end-to-end feature ownership
• 🎨 Frontend Engineering — responsive, modern UIs with React, Next.js & Tailwind CSS; performance-focused and mobile-first
• 🏗️ Backend & Architecture — multi-role JWT auth, SaaS products, scheduled jobs, payment integrations, serverless & database design
• 🤖 AI Integration — building products on top of LLM APIs (Groq, Claude) with streaming, RAG, and real-time interaction
• 🚀 SaaS Product Development — shipped on a live production platform with real users, real payments and real deadlines

Languages: TypeScript · JavaScript · Python · SQL · Java (learning DSA) Frontend: React · Next.js · Tailwind CSS · React Query (TanStack) Backend: Node.js · Express · REST APIs · JWT Auth · node-cron Databases: PostgreSQL · MongoDB · Prisma ORM · MongoDB Atlas AI & Tools: Groq · Anthropic (Claude) API · Razorpay · Git · Vercel · Render · Postman

Full-stack across the web app, mobile app and backend of a live sports-venue booking product — live on the web, Google Play and the App Store. I owned modules end to end, and left with an appreciation letter from the team.

What I shipped:

• 🧾 Core booking engine — guest booking, multi-slot booking and Razorpay payment integration, with webhook-confirmed bookings and refunds
• 💸 Split payments — several players pay for one booking through shared payment links, with payment holds and automatic refunds on expiry
• 🗂️ Owner, Manager and Admin portals — slot and pricing management, payouts, coupons, approval workflows and analytics
• 📅 Schedule templates & dynamic pricing — recurring weekly schedules with per-date overrides; peak/off-peak rules calculated timezone-safe
• 🏃 Running-club module — paid event registration and GPS-based run tracking
• 🔐 Auth & notifications — email OTP and Google OAuth; scheduled reminders, WhatsApp booking confirmations and in-app alerts
• ⚡ Backend performance — pagination, caching, database indexing and query optimisation, including eliminating N+1 queries and fixing a timezone bug on slot boundaries

The lesson that stuck: code behaves differently once real users and real money move through it — payment states that fail halfway, edge cases you never hit locally, and performance problems that only appear at scale.

Next.js 15 · React 19 · Express · Prisma · PostgreSQL · Razorpay · TanStack Query

sequenceDiagram
    autonumber
    participant P as 🧑 Player
    participant W as Next.js App
    participant A as Express API
    participant D as PostgreSQL
    participant R as Razorpay
    participant N as Notifications

    P->>W: Pick venue + multiple slots
    W->>A: POST /bookings/checkout
    A->>D: Validate schedule + price rules
    A->>D: Reserve slots in a transaction
    A->>R: Create payment order
    R-->>P: Payment sheet
    P->>R: Pay
    R-->>A: Webhook payment.captured
    A->>D: Confirm booking + write ledger
    A->>N: Fan out confirmations
    N-->>P: WhatsApp + in-app alert
    A-->>W: Booking confirmed

A scam guardian for the people UPI brought online and nobody built for. It flags a fraud before the money moves, explains which trick was used in plain Gujarati, and keeps working with no internet at all — because the detector itself ships to the phone.

The part I care about most: the AI never decides what is a scam. Rules plus a LightGBM model score every known tactic and set the verdict; the local model only translates that verdict into her words, with the actual RBI/NPCI advisory attached. That's why a small offline model can be trusted here.

• 📴 Offline-first — real engine exported to the device; verdict in ~150 ms, nothing uploaded
• 🗣️ Gujarati-first — language is the first screen, every verdict can be spoken, elder mode throughout
• 🎧 Practice a scam call, safely — the app plays the fraudster in a real Gujarati voice and coaches her through spotting each trick
• 📸 Four inputs, one engine — SMS/WhatsApp text, links, screenshots (the fake "collect request" trap) and voice notes
• ☎️ One tap to act — ask a family member, or report to the 1930 cybercrime helpline

Next.js PWA · FastAPI · Rules + LightGBM · Qwen (local) + Groq · Chroma · Whisper + edge-tts · Docker

Built for the Drift-Sense problem statement (Applied Materials): a wafer inspection tool must return to the same die site thousands of times a day, but the stage drifts — and the layout repeats everywhere, so template matching hands back hundreds of near-identical matches.

Our answer is deliberately classical CV, grounded in SEM physics — no deep learning, nothing in the judges' as-is run that can break. We generate DRAM-style data with real SEM noise (Poisson + Gaussian, edge brightening), then localize with a multi-scale ZNCC sweep, the official centre rule, and a sub-pixel fit. Output is (x, y) plus a PSR confidence — when a field is genuinely ambiguous, we say so instead of guessing.

• 📐 Incommensurate mat pitches were the breakthrough — pure lattice data ranked the truth ~762nd under noise; landmarks that never line up inside the frame took it to rank 0, 0.43 px error
• ⚡ CPU only — no GPU, no model weights, under a second per pair
• 🧾 Every generator constant is citation-tagged in code [S1]–[S12], traced to SEM physics or DRAM 6F² literature
• 🎯 Honest failure — the PSR flag exists for the deliberately ambiguous periodic region in the official test set

Python 3.10 · NumPy · OpenCV · SciPy · ZNCC multi-scale matching · Matplotlib

Team DriftLock — Het Patel (lead) · Eklavya Jha, Gandhinagar University

My most-worked project of 2026. A centralised CCTV registry with GIS and vehicle analytics that reconstructs a vehicle's route across cameras, constrained by real geography rather than guesswork — watchlist alerts in under 50 ms, forensic search in under 150 ms, running fully offline inside a police network.

The part I'm proudest of isn't a feature: we built it on measurements taken from the organisers' own camera feeds, and the submission states plainly why ANPR cannot work at their camera geometry instead of demoing a number we couldn't hit in the field.

GIS · computer vision · cross-camera re-identification · offline deployment

Anomaly, fraud and inefficiency detection across MPLADS public-spending data — the Ministry of Statistics problem statement (PS 26102) at Smart India Hackathon 2026, built with Team Technoverse.

anomaly detection · fraud analytics · public-spending data

A planner where the route, the calendar and the money are the same object — change the itinerary and the budget and schedule follow, instead of drifting out of sync across three separate screens.

Next.js 16 · React 19 · PostgreSQL · Prisma 6

A drop-in reimplementation of json-server using only the Node standard library — no dependency tree at all — with live WebSocket updates and a built-in dashboard on top.

Node stdlib only · WebSocket · REST

Turns phone captures into finished work with a local LLM running in the browser — nothing leaves the device, and it lands on your desk.

WebLLM · Next.js · React Three Fiber

An AI interviewer that spans 21 tech roles with adaptive, real-time questioning in voice or text mode. Each ~30-minute session ends with per-question scored feedback and a personalized improvement plan. Features a terminal-style UI with token-streamed responses.

Next.js · TypeScript · Groq (Llama 3.3 70B) · Web Speech API

flowchart LR
    U["🎙️ Candidate"] -->|voice or text| UI["Terminal UI<br/>Next.js + TypeScript"]
    UI -->|Web Speech API| STT["Speech to text"]
    STT --> API["Interview route<br/>edge runtime"]
    UI --> API
    API --> CTX["Session context<br/>role + difficulty + history"]
    CTX --> LLM["Groq · Llama 3.3 70B"]
    LLM -->|token stream| API
    API -->|SSE| UI
    CTX --> ADAPT["Adaptive engine<br/>21 role tracks"]
    ADAPT --> SCORE["Per-question scoring"]
    SCORE --> PLAN["📊 Improvement plan"]

    classDef c fill:#8b5cf6,stroke:#c4b5fd,color:#fff,stroke-width:1px
    classDef a fill:#0ea5e9,stroke:#7dd3fc,color:#fff,stroke-width:1px
    classDef o fill:#a855f7,stroke:#e9d5ff,color:#fff,stroke-width:1px
    class U,UI,STT c
    class API,CTX,LLM,ADAPT a
    class SCORE,PLAN o

Transforms any GitHub user's commits, repositories, and languages into an AI-generated, magazine-style developer narrative in seconds, with shareable editorial output.

Next.js 15 · MongoDB · Anthropic (Claude) API

flowchart LR
    IN["👤 GitHub username"] --> GH["GitHub REST + GraphQL<br/>commits · repos · languages"]
    GH --> AGG["Aggregation layer<br/>signals + timeline"]
    AGG --> CACHE[("MongoDB<br/>story cache")]
    AGG --> CLAUDE["Anthropic Claude<br/>narrative generation"]
    CLAUDE --> STORY["📰 Magazine-style story"]
    CACHE --> STORY
    STORY --> SHARE["Shareable page"]

    classDef c fill:#8b5cf6,stroke:#c4b5fd,color:#fff,stroke-width:1px
    classDef a fill:#0ea5e9,stroke:#7dd3fc,color:#fff,stroke-width:1px
    classDef o fill:#a855f7,stroke:#e9d5ff,color:#fff,stroke-width:1px
    class IN,GH c
    class AGG,CACHE,CLAUDE a
    class STORY,SHARE o

Seven real businesses running on sites I designed, built and deployed. Each one is a paying client's public front door, not a portfolio piece.

One complete build per industry, made so a prospective client can see their own business before committing. Layout, copy, responsive pass and deploy — all of it real.

→ See all 19+ projects at buildbyhet.me

python3 assets/generate-3d-assets.py
python3 assets/generate-stats-card.py Het161

💡 "Ship it. Iterate. Ship again."

Open to full-time & internship roles · Ahmedabad, India 🇮🇳