Hello from Kanpur: what I build, and what I'll write about here

I'm Pranjul Rathour, a GenAI engineer from Kanpur, India. This is the first post on this blog, so here is who is writing, what I have actually shipped, and what you can expect if you follow along.

What I build

Between February and July 2026, at Next Upgrad Web Solutions, I designed, built and shipped five production AI applications end to end, alone, and deployed all of them with Docker, CI and observability on Railway and Vercel:

• RAG.NextUpgrad: a retrieval-augmented generation service with hybrid dense plus BM25 retrieval, cross-encoder reranking, streaming cited answers, and a confidence gate that refuses to call the model when the evidence is weak. It runs in about 220 MB of memory, inside a free 512 MB instance.
• FineTune Studio: fine-tune open models with QLoRA in a browser. Upload a dataset, validate it, train, watch the loss stream live, then compare base and tuned models side by side. Trains Qwen3-1.7B at about 3.2 GB of VRAM. 107 tests pass.
• FaceVision: face detection, recognition and liveness checks that run entirely in the browser with ONNX Runtime Web. The backend only ever stores embeddings, never images.
• DocuLens AI: document intelligence for invoices and contracts, with extraction, comparison and deterministic compliance scoring.
• OCR & Speech Workspace: page-batched PDF OCR, live speech-to-text and document-scoped chat, built on Mistral's models.

Before that, I co-founded SCULT INDIA, where I lead engineering for a 14-member team, built the internal Studio CRM the agency runs on, and shipped 5+ paid client products. I'm finishing a BCA at Dr. Virendra Swarup Institute of Computer Studies, Kanpur.

What I have won

• 1st prize, Changethon 2025, National Social Summit, IIT Roorkee, with Krishi Gyan, an AI agribot for farmers.
• 1st prize, Product Genesis, Vividhotsava 2025 at CSJMU Kanpur, for BrandHive, presented by the Vice Chancellor.
• 1st prize, BYTEBATTLE by MeetKats, with Annapurna, a food-sharing platform for donors, NGOs and volunteers, against 15+ teams.
• Top 3 at IIT Kanpur's hackathon.

None of those came from knowing the most algorithms. They came from scoping tightly, shipping a working deploy before the demo, and explaining one tradeoff clearly. I'll write about that.

Who I mentor

I founded TechVerse Enclave, a developer community of 500+ members, and have mentored 200+ students in web development, AI and careers. In 2025 I ran two sessions at my own college, VSICS Kanpur: an AI-powered software development session in April, and the BCA induction in July, where I gave 100+ first-year students a three-year roadmap from first project to placement or startup.

What this blog will cover

One long-form piece a week, always from something I actually built or ran:

1. Production RAG that refuses to hallucinate: retrieval, reranking, confidence gates, citations.
2. Fine-tuning open models on a student budget: QLoRA, dataset validation, honest evals.
3. Vision, OCR and speech in real products: privacy-by-design, browser-side inference.
4. Shipping AI as a small team: the unglamorous parts that keep a service up.
5. From college hackathons to production: what wins, what gets you hired, portfolio over certificates.
6. Running a student tech community without burning out.

Short daily takes go out on LinkedIn and Bluesky. Code lives on GitHub.

Invite me

I run free hands-on sessions for college tech clubs and placement cells, and I'm open to judging hackathons, mentorship sessions and guest talks at IITs, NITs and any campus that wants its students shipping real things. Write to pranjulrathour41@gmail.com or message me on LinkedIn.

If you're a student reading this: pick one problem you have, ship the ugliest working version this week, and write 500 honest words about what broke. That loop is the whole method. See you next week.