Rohan Kumar

01 / Selected work

Systems that ship.

Four case studies from real repositories — every link verified, every metric taken from source. Open a case study for architecture and decisions.

01

AI Agents · LLM Systems

Cortex Agent

An autonomous multi-model AI agent studio, built on NVIDIA NIM frontier models.

Problem: Hosted chat UIs cap context, hide tool use, and trap artifacts inside the conversation.
Solution: An autonomous agent studio — a ReAct loop running real tools, streamed into a Claude-style artifacts canvas.
Engineering: FastAPI gateway over NVIDIA NIM (1M-token context), 7 tools, SQLite WAL, SSE heartbeat multi-tab sync, PBKDF2 + HMAC auth, daily token quotas.
Result: Deployed at cortex-agent-tawny.vercel.app with automated test suites covering auth, tools and sync.

1M

Max context window

128K

Max output tokens

7

Autonomous tools

FastAPI
LangChain
NVIDIA NIM
SQLite WAL
SSE
Vanilla JS

Live deployment GitHub

Cortex Agent landing page — autonomous multi-model AI agent studio

NostalgiaNet++ — a digital time capsule for the moments that matter

02

Product · Full-Stack

NostalgiaNet++

A digital time capsule — seal memories now, unlock them years from now.

Problem: Memory apps are photo dumps. A time capsule's whole point is a lock enforced by software — not hidden with CSS.
Solution: Seal photos, videos and letters inside TimeVaults with future unlock dates — group invites pre-seal, public discovery, unlock-day notifications.
Engineering: Next.js App Router + Prisma/PostgreSQL (Neon), NextAuth (bcrypt + OAuth), byte-signature upload validation, server-side sealed-content redaction on every route.
Result: Live at nostalgia-net-silk.vercel.app — runs entirely on free tiers, $0/month infrastructure.

Next.js
TypeScript
Prisma
PostgreSQL
NextAuth
Vercel Blob

Live deployment GitHub

03

Deep Learning · IIT Madras

Smart MCQ Solver

A 4-model ensemble for 5-option multiple-choice QA — scored by MAP@3.

Problem: 5-option MCQ ranking scored by MAP@3 — while duplicated questions pushed naive local validation to a perfect ~1.0.
Solution: Four deliberately diverse models — TF-IDF+LogReg, a from-scratch BiLSTM, fine-tuned DeBERTa-v3-small and RoBERTa-base — fused by weighted rank voting.
Engineering: Leakage audit (242 duplicate train rows, 267 train/test overlaps), W&B-tracked training, clean src/ pipeline, Gradio demo on Hugging Face Spaces.

PyTorch
Transformers
scikit-learn
Weights & Biases
Gradio

Live deployment GitHub

Result — verified leaderboard · MAP@3

TF-IDF + LogReg0.0000
BiLSTM (from scratch)0.0000
DeBERTa-v3-small0.0000
RoBERTa-base0.0000
Weighted ensemble0.0000

Verified from the repository README. Bar scale runs 0.70 → 0.77 to make differences visible. Ensemble beats every single model.

04

Examination Management Portal

Backend Systems · Flask

Institution-grade exam lifecycle with zero client-side JavaScript.

Problem: Exam administration fails at the edges: double-booked seats, vague grading criteria, unaudited admin overrides.
Solution: A 3-role exam lifecycle platform — RBAC with approval gates, dynamic rubric grading, concurrency-safe slot booking.
Engineering: 100% server-rendered PRG (zero client JS), atomic seat decrement with seat recovery, schema-level constraints, BookingHistory audit log, REST /api/v1.
Result: 11/11 automated tests passing — auth, RBAC isolation, concurrency, cancellation, grading.

Flask 3.1
SQLAlchemy 2.0
SQLite
Bootstrap 5
Jinja2

GitHub

Additional projects

smaller scope, same standards

Competitions

Smart MCQ Solver Challenge: pick the top 3 of 5 options, scored by MAP@3.

Model	Approach	MAP@3
TF-IDF + Logistic Regression	Baseline	0.7511
BiLSTM (from scratch)	Sequence model	0.7540
DeBERTa-v3-small	Fine-tuned transformer	0.7544
RoBERTa-base	Fine-tuned transformer	0.7544
Weighted rank-vote ensemble	All four combined	0.7602

02 / About

I don't stop at notebooks.

I'm a third-year B.S. Data Science & Applications student at IIT Madras, and I spend most of my time doing the same thing: taking an idea from a rough problem statement to a system that actually runs — designed, implemented, evaluated, and deployed. Each project has real architecture, tests, and a live URL where possible.

My work sits where machine learning meets software engineering. On the ML side that means building models the honest way — baselines first, leakage audits, leaderboard-anchored evaluation — like my Smart MCQ Solver, where a four-model ensemble beat every individual model. On the engineering side it means agents with real tool loops, concurrency-safe booking systems, JWT stateless APIs, and background job pipelines.

I'm drawn to AI/ML engineering and full-stack development for the same reason: both reward fundamentals over hype. I like free-tier deployments, boring databases, and code someone else can actually read. Currently deepening applied ML — and taking every project from notebook to a system that actually runs.

Stack I use

Languages, frameworks and tools across ML pipelines and full-stack apps.

Python
PyTorch
scikit-learn
FastAPI
Flask
Next.js
React
Vue.js
TypeScript
PostgreSQL
SQLite
Redis
Tailwind
Git
Linux
Python
PyTorch
scikit-learn
FastAPI
Flask
Next.js
React
Vue.js
TypeScript
PostgreSQL
SQLite
Redis
Tailwind
Git
Linux

04 / Engineering approach

Idea to deployment, on purpose.

I like building end-to-end systems rather than isolated notebooks — the same loop, every time, from framing to a live URL.

01
Idea
Interrogate the actual constraint before touching code. What breaks today, for whom, and what would 'solved' measurably mean?
02
System design
Design the data model and boundaries first — schemas, state machines, contracts. Most late bugs are early decisions.
03
Implementation
Build the thinnest honest version end to end: real auth, real persistence, real errors. No mock stages that never ship.
04
Evaluation
Tests for the edges that matter — concurrency, authz boundaries, leakage audits. Leaderboards over flattering local metrics.
05
Deployment
Ship it publicly, wire the cron, smoke-test the route. A system isn't finished until it's reachable.

“This person doesn't just learn technologies — they actually build systems.”
The bar I build against

05 / Code & experiments

Public by default.

Selected repositories

github.com/23f2003236

Full history — including coursework, notes and experiments — on the profile.

Kaggle

lukerohan

Smart MCQ Solver Challenge

The IIT Madras Deep Learning course ran as a Kaggle-style competition: 5-option MCQs, scored by MAP@3. The final weighted ensemble of four models scored 0.7602 MAP@3 on the competition leaderboard — above every individual model — with the submission notebook published in the repository.

$ honest-ml note — local validation was invalidated by data leakage (242 dup rows, 267 train/test overlaps). Leaderboard was the only trusted signal.

06 / Résumé

One page. Zero fluff.

An ATS-friendly, single-page PDF generated from the same verified facts as this site — education, stack, four deeply documented projects, and a compressed view of the rest. No invented experience, no padding.

Download résumé (PDF)Request a tailored copy

Inside the PDF

01Professional summary
02Education — IIT Madras
03Technical skills — 6 groups
044 selected projects with verified outcomes
05Additional projects — 5, compressed

format: A4 · serif · selectable text · print-ready

Command Palette

Systems that ship.

Cortex Agent

NostalgiaNet++

Smart MCQ Solver

Examination Management Portal

Competitions

I don't stop at notebooks.

Stack I use

Idea to deployment, on purpose.

Idea

System design

Implementation

Evaluation

Deployment

Public by default.

Selected repositories

Kaggle

One page. Zero fluff.

Let's build something worth deploying.