Manswi Pradhan

MANSWI PRADHAN

I build tools that explain themselves — a Caesar-cipher cracker that shows its chi-squared math, a medical chatbot that shows its reasoning. Currently deep in CTF challenges and web security.

The thread connecting it: explainable systems — whether the domain is cybersecurity, applied ML, or the math underneath both.

Things I currently believe

Second-year Mathematics & Computing at IIIT Raichur, arrived with a Mass Communication background instead of a coding one. Here's what that combination has taught me so far — subject to revision as evidence comes in.

on ciphersEvery complex cipher is just statistics wearing a disguise.

on toolsAn answer without the "why" is just a guess with better formatting.

on backgroundA Mass Comm background isn't a handicap in a math degree — it's a debugging tool for other people's confusion.

on bugsMost interesting bugs are just wrong assumptions with excellent uptime.

on CTFsCTFs are 50% cryptography, 50% patience, and 100% snacks.

on differentiationIn a peer group that's very ML-heavy, cybersecurity turned out to be a fine hill to plant a flag on.

Research notes

CTF Learning Assistant

Solo project

Question: Most CTF tools just hand you an answer — what if one taught the underlying concept first?
Approach: Built browser-based detectors for 10+ encodings (Base64, Hex, Binary, ROT13, Vigenère, Morse, Octal, A1Z26, URL, XOR/Base32).
Discovery: Chi-squared frequency analysis can guess a Caesar shift without brute-forcing all 26 — letters gossip about themselves.
Improve: Extend the frequency analysis so it holds up on short ciphertexts, where the statistics get shaky.

JavaScriptHTML/CSSCybersecurity

live demo → source →

MediFind

3-person team · supervised by Dr. Kaushiki Roy

Question: Can a chatbot triage symptoms and point someone to the right specialist, not just "a doctor"?
Approach: Owned the full frontend, a semantic specialist-matching layer using Sentence Transformers, and symptom extraction/classification via Groq/LLaMA + FastAPI, backed by a keyword fallback.
Discovery: "AI-powered" search still needs a boring backup plan for when the vectors don't cooperate.
Improve: Tighten the fallback logic so it reads less like a backup and more like a second opinion.

PythonFastAPISentence TransformersGroq/LLaMAMySQL

source →

Fruit Recognition — VGG16 Transfer Learning

3-person team

Question: Can a model tell 15 kinds of fruit apart — and does the choice of loss function actually matter?
Approach: Transfer-learned VGG16 backbone (frozen conv layers, custom classifier head), comparing CrossEntropy vs Focal Loss, plus the training loop.
Discovery: Best experiment hit 97% test accuracy — turns out fruit is more classifiable than people.
Improve: Test against messier real-world photos instead of clean dataset images.

PythonPyTorchTransfer Learning

source →

picoCTF Writeups

Solo

Question: How do you document a CTF solve so it's still useful to someone who isn't you in six months?
Approach: Organized writeups by category (Forensics, Cryptography, General Skills), each covering the tools used and the dead ends hit.
Discovery: The dead ends usually teach more than the solve does.
Improve: Add a "time to solve" note per writeup to track improvement over time.

CryptographyForensicsCTF

source →

Skills, plotted instead of bar-charted

Progress bars don't say much. This is confidence against hours actually logged — hover or tab to a point for the honest version.

Python Hover or tab to a point on the plot to read the honest version of that skill.

CodeChef (newton_lebniz) — Rookie league, 882 rating · 1246 DSA rating · 110+ problems solved across 8 rated contests.

From the notebook

Short writeups on the math and reasoning behind what I build — not just what worked, but why.

Building a chi-squared Caesar cipher cracker read →

How letter-frequency statistics let a script guess the shift key without brute-forcing all 26 — and where the math breaks on short ciphertexts.

A picoCTF cryptography writeup, properly explaineddraft

Walking through one challenge start to finish — the wrong turns included, not just the solution.

What building MediFind's vector search taught me about semantic matchingdraft

Sentence Transformers, keyword fallback, and why "AI-powered" search still needs a boring backup plan.

Things I currently believe

Research notes

CTF Learning Assistant

MediFind

Fruit Recognition — VGG16 Transfer Learning

picoCTF Writeups

Skills, plotted instead of bar-charted

GitHub

From the notebook

If you've read this far, we probably have something to talk about.