DEV Community: PRANJUL RATHOUR

Choosing an LLM provider: a decision framework

PRANJUL RATHOUR — Fri, 25 Sep 2026 04:32:26 +0000

New frontier models ship every few weeks now. Chasing the newest one is a losing game unless your architecture makes switching free.

Put the provider behind one interface

In every NextUpgrad app, the LLM sits behind one port, and swapping Groq for Anthropic or OpenAI is a config change, not a rewrite. When a new model ships, changing one line and rerunning the evals is the entire migration.

What actually differs between providers

Latency: Groq's inference speed matters when a user is waiting on a streamed answer.
Context window: matters for document-heavy tasks like OCR chat, less for short classification.
Cost per token: matters at scale, less for a low-traffic student project.
Tool-use reliability: matters enormously for agentic workflows, barely at all for single-turn Q&A.

The evaluation that actually predicts your bill

A public benchmark tells you who won an exam. Your own evals, on your own documents and your own budget, tell you who you'd actually hand production traffic to.

Pranjul Rathour · GenAI engineer, 3x hackathon winner, campus mentor. Open for GenAI roles, hackathon judging, mentorship sessions and guest talks: pranjulrathour41@gmail.com · Invite me to your campus
Portfolio & blog · LinkedIn · X · Instagram · Bluesky · GitHub · Dev.to

How to Invite a GenAI Speaker to Your College

PRANJUL RATHOUR — Thu, 24 Sep 2026 13:34:22 +0000

How to invite a GenAI speaker to your college — the whole checklist.

The email is two lines: two dates + your students' year. Pick a talk (RAG, fine-tuning on a free GPU, shipping AI as a small team, what wins a hackathon, BCA to GenAI engineer), a 90-minute workshop, or a judging seat. The AV list: HDMI tested the day before, power for laptops, Wi-Fi for 60, one coordinator who replies.

Free for student communities. Previous: VSICS Kanpur (talk show, mentorship session, BCA induction for 100+), IIT Roorkee, CSJMU, MeetKats.

Pranjul Rathour · GenAI engineer, Kanpur · pranjulrathour41@gmail.com · pranjulrathour.scult.in/invite

GuestSpeaker #CampusTalk #Hackathon #StudentCommunity #GenAI

Three First Prizes, One Loss

PRANJUL RATHOUR — Wed, 23 Sep 2026 13:35:19 +0000

Three first prizes, one year, one loss that taught more.

Feb — Changethon, IIT Roorkee: Krishi Gyan, an AI agribot (React, Flask, Gemini Flash).
Apr — Product Genesis, CSJMU: BrandHive, a pitch.
Jun — BYTEBATTLE by MeetKats: Annapurna, surplus-food coordination (React, Supabase), vs 15+ teams.
Dec — STROT, 48 hours, didn't win. Lesson: clarity and storytelling matter as much as the system.

Same team every time: Raman Shukla, Saksham Gupta and me. Open to judge hackathons and give this as a talk, free for student communities.

Pranjul Rathour · GenAI engineer, Kanpur · pranjulrathour41@gmail.com · pranjulrathour.scult.in/invite

Hackathon #StudentDeveloper #BuildInPublic #GenAI #Kanpur

1,211 Prompts, Every One Stamped

PRANJUL RATHOUR — Tue, 22 Sep 2026 13:32:42 +0000

1,211 prompts, every one stamped with the model and date it was tested on.

52 categories in 9 groups. Worth your time first: Claude Code (42), AI Agents & RAG (19), Veo (34), Nano Banana (30), Build Apps Without Code (23), LinkedIn (32). Every prompt page shows the verification stamp, fillable variables and a why-it-works note.

Browse: tools.scult.in/prompts · from your agent: tools.scult.in/mcp

Built solo by Pranjul Rathour, GenAI engineer, Kanpur · pranjulrathour41@gmail.com

PromptEngineering #ClaudeCode #GenAI #AITools #BuildInPublic

7 Levels of RAG Apps

PRANJUL RATHOUR — Mon, 21 Sep 2026 13:32:37 +0000

7 levels of RAG apps — which level is yours at?

1 Notebook · 2 Chunker · 3 Hybrid (dense + BM25) · 4 Re-ranker · 5 Gatekeeper (a confidence gate that refuses to answer) · 6 Auditor (citations + 196 tests) · 7 Operator (multi-provider fallback, ~220 MB in prod).

Level 7 is RAG.NextUpgrad, which I built and open-sourced: github.com/Pranjulrathour/RAG.NEXTUPGRAD · write-up at pranjulrathour.scult.in/projects/rag-nextupgrad

I teach this as a hands-on college workshop, free for student communities. Pranjul Rathour · GenAI engineer, Kanpur · pranjulrathour41@gmail.com

RAG #LLM #GenAI #AIEngineering #Python

5 Production AI Apps, All Open Source

PRANJUL RATHOUR — Sun, 20 Sep 2026 13:32:29 +0000

5 production AI apps, all open source, with the real numbers.

RAG.NextUpgrad — hybrid retrieval, a confidence gate that refuses to answer, 196 tests.
FineTune Studio — QLoRA in a browser, live loss, 107 tests.
FaceVision — face recognition that never uploads a face, 260+ tests.
DocuLens AI — deterministic document scoring, 44 tests.
OCR & Speech Workspace — 1,500-page PDFs, live speech, 40 tests.

All repos: github.com/Pranjulrathour · write-ups: pranjulrathour.scult.in/projects
Pranjul Rathour · GenAI engineer, Kanpur · pranjulrathour41@gmail.com

GenAI #RAG #OpenSource #MachineLearning #BuildInPublic

15 Free Tools, No Signup, Every Link

PRANJUL RATHOUR — Sat, 19 Sep 2026 13:32:35 +0000

15 free tools, no signup, every link inside.

SEO: schema + FAQ generators, UTM builder, ROI calculator, speed test. Business: invoices (GST/VAT, 8 currencies), names, slogans, email signatures. Dev/design: JSON, QR (with UPI), favicons, word counter, OKLCH palettes. Plus the AI Visibility Checker — can ChatGPT even see your site?

13 of 15 never upload a byte. All of it also reachable from your AI agent via the MCP endpoint (slide 8).

tools.scult.in — built solo by Pranjul Rathour, GenAI engineer, Kanpur · pranjulrathour41@gmail.com

FreeTools #SEO #WebDev #IndieHacker #BuildInPublic

15 Free Tools, No Signup, Every Link

PRANJUL RATHOUR — Sat, 19 Sep 2026 04:24:19 +0000

15 free tools, no signup, every link inside.

13 of 15 never upload a byte. All of it also reachable from your AI agent via the MCP endpoint (slide 8).

tools.scult.in — built solo by Pranjul Rathour, GenAI engineer, Kanpur · pranjulrathour41@gmail.com

FreeTools #SEO #WebDev #IndieHacker #BuildInPublic

Building your personal brand as a CS student

PRANJUL RATHOUR — Fri, 18 Sep 2026 04:32:31 +0000

I tell every incoming batch the same thing: LinkedIn, YouTube and Instagram are your resume now. Nobody's waiting for permission to check your online presence before deciding whether to take a meeting.

The version that doesn't feel fake

Post the build, not a highlight reel. A short update on a project you're mid-way through, including the part that's not working yet, reads as more credible than a polished announcement — and it's easier to actually write, because you don't have to invent confidence you don't have yet.

Consistency beats production value

A rough weekly post beats a polished monthly one. The algorithm and the audience both reward showing up, not the editing quality of any single piece.

GitHub is not optional

Whatever platform you post updates to, the actual proof lives on GitHub. An empty profile says something too — usually the wrong thing.

Three times I didn't win first place (and what stuck more than the wins)

PRANJUL RATHOUR — Thu, 17 Sep 2026 04:33:37 +0000

My hackathon record has three first prizes in it, but the finishes that actually changed how I build are the ones where I didn't win outright.

Top 3 at IIT Kanpur

A top-3 finish at an IIT Kanpur hackathon sits right next to my three wins on the resume, and it taught me the same lesson every near-miss does: the gap between first and third is rarely the code. It's almost always the last ten minutes of the pitch.

1st Runner-Up, Cyber Security Awareness Event

At VSICS Kanpur's Cyber Security Awareness Event, my teammate Ujjwal Mishra and I placed 1st runner-up. The event deepened my understanding of real threats and preventive measures — the kind of foundational security thinking that shows up years later in decisions like FaceVision's embeddings-only backend.

2nd Runner-Up, Perspectives Unleashed

In a debate competition organised by the Computer Applications Department at VSICS, I placed 2nd runner-up arguing a position I didn't personally hold. Every rebuttal sharpened my ability to think on my feet in front of a room — a skill that has mattered more in client pitches and mentorship sessions than any algorithm I've written.

Why near-misses are worth writing about

A trophy tells you what worked. A near-miss tells you exactly what didn't, in specific and uncomfortable detail. If you're a student and you just lost a competition you thought you'd win, write down the specific moment it slipped, before the memory softens it. That note is worth more than the certificate would have been.

NIMCET, placements, or a startup: choosing your Year 3

PRANJUL RATHOUR — Wed, 16 Sep 2026 04:32:30 +0000

By Year 3, most BCA students are quietly panicking about three paths at once: higher studies through NIMCET or MCA, campus placements, or trying to build something of their own. Splitting focus across all three usually means being mediocre at each.

NIMCET / MCA: go all in, or don't go at all

This path rewards depth in a narrow set of topics over a fixed timeline. Half-preparing for an entrance exam while also job-hunting is the worst version of both strategies.

Placements: skills plus proof, not just a resume line

A portfolio with real, deployed projects changes a placement interview from a quiz about theory into a conversation about decisions you actually made. That conversation is easier to win.

Startup: build the MVP, form the team, before you decide

I co-founded SCULT INDIA out of exactly this path, alongside teammates I'd already built with in hackathons. The team and the MVP should exist before the decision to go all-in does, not after.

The one rule that applies to all three

Whichever you choose, mentor a junior along the way. It's the step everyone skips and the one that compounds the most — it completes the circle instead of leaving it open.

How to start a tech community on your campus

PRANJUL RATHOUR — Tue, 15 Sep 2026 04:32:26 +0000

TechVerse Enclave started in 2023 with a handful of people who actually wanted to build things together, not a launch announcement to hundreds of strangers. That order matters more than most people starting a community realise.

Start with five, not fifty

A small founding group that actually shows up beats a large one that doesn't. Five committed admins running consistent sessions will outgrow fifty passive members within a semester.

Pick one format and prove it works

Live-build sessions with a real project on screen consistently outperformed slide-only talks in our community. Find your version of that before trying to run five different event types at once.

Let momentum do the recruiting

We didn't scale TechVerse Enclave to 500+ members by advertising — we scaled it by running sessions worth talking about, and mentoring people well enough that they brought others.

The metric that matters, and the one that doesn't

Member count is vanity until it converts to mentored students actually shipping things. 200+ students mentored through TechVerse Enclave matters more to me than the 500+ member count that gets quoted alongside it.