AI Engineer & Full-Stack Developer · AI Engineer at EnxtAI · Noida, India
I build production machine-learning pipelines and the backends that carry them: computer vision models that read lips and count crowds, scraper clusters that ingest ten thousand pages a minute, and webhook gateways that answer inside Meta’s 200 ms window.
B.Tech CSE, Honors in AI & ML — Amity University (2021–2025)
Speech recognition fails the people who most need captions. LumaVoice reads speech from the mouth alone, so a silent video still says something.
A CNN-LSTM spatio-temporal network in PyTorch that classifies phonemes from the mouth region at 30 FPS, behind a FastAPI service and a React front end that takes uploads or a live webcam.
Built with PyTorch, FastAPI, React 18, TypeScript, Docker
App repo Model notebook Aligner serviceCCTV operators need two answers at once: how many people are here, and is something going wrong. Single-purpose models give one or the other.
One checkpoint that bundles a DINOv3 density counter, a V-JEPA motion encoder and a Qwen2.5-VL reasoner. Every change ships only if it beats criteria I register before running the evaluation.
Built with PyTorch, DINOv3, V-JEPA, Qwen2.5-VL, vLLM, Docker, Modal
NWPU-Crowd mirrorFan-out scrapers either flood the target or drown in duplicates. Throughput is only useful if every page lands exactly once.
A back-pressured fan-out cluster on a RabbitMQ topic exchange, with Redis SHA-256 fingerprints so ingestion stays duplicate-free. Measured at 10,480 pages per minute.
Built with Python, RabbitMQ, Redis, Docker
Meta’s Cloud API drops webhooks that are not acknowledged within 200 ms, and a hospital cannot lose a patient’s message.
An Express webhook that verifies HMAC-SHA256 signatures, acknowledges in 142 ms, then handles booking and encrypted report delivery asynchronously.
Built with Node.js, Express, Meta Cloud API, Queues
Deal news in consumer goods is scattered across hundreds of outlets and repeated endlessly. Analysts need the deals, once, with context.
A pipeline of agents: news ingestion from Tavily, event extraction with Gemini, relevance and duplicate filtering over pgvector, and a writer agent that turns recent deals into a newsletter.
Built with TypeScript, Tavily, Gemini, PostgreSQL, pgvector
RepoThe court portal has no API or bulk download: 30-day query windows, a maths CAPTCHA on every search, and results injected by AJAX. A naive crawl needs about 27,000 page loads.
An automated scraper that plans the date windows, answers the arithmetic CAPTCHA, parses the AJAX responses and collects every public judgement with a results report.
Built with Python, HTML parsing
Repo ResultsStudents, instructors and admins need different views of the same courses, grades and deadlines, and none of them want to hunt for what is due next.
An ASP.NET Core 8 API with Entity Framework and cookie auth, and an Angular 19 front end: enrolment, assignments, a SpeedGrader-style grading screen, a gradebook with CSV export and one-tap demos.
Built with ASP.NET Core 8, EF Core, Angular 19, TypeScript
RepoOne chatbot voice does not fit every moment. Advice from a mentor, patience from a listener and a yoga cue are different conversations.
A mobile app where you talk to specialised personas (mentor, teacher, motivator, listener, philosopher, yoga instructor) or define your own.
Built with TypeScript, Mobile, LLM APIs
RepoRace weekends scatter across timing pages, standings and highlight clips. I wanted one screen for all of it.
A Next.js dashboard with race history charts and a YouTube highlights section, deployed on Vercel.
Built with Next.js, TypeScript, Vercel
Live site RepoReports and decks are repetitive to produce, and work done by agents on a laptop leaves no audit trail.
A multi-agent system that runs inside GitHub Actions, plans its own tasks and commits version-controlled DOCX, PPTX and structured data artefacts.
Built with Python, GitHub Actions, LLM agents