VIT Vellore, India CSE student · Builder · Open source

Hello, I’m Vaishnavi

I build useful things for real problems.

I’m a Computer Science student at VIT Vellore building across AI/ML, computer vision, full-stack products, and developer tools. I like turning a rough idea into something people can actually use — and explain.

See my projects Read my resume

Open to learning & building ( Scroll )

01A little context

Curious about how things work and how to make them better.

I’m a CSE student at VIT Vellore, and I learn best by making. My projects move between AI systems, computer vision, web products, and tools that help people learn, work, or respond to real problems.

I enjoy taking a project from the first sketch to a working version, then making it clearer through feedback. Readable code, simple interfaces, and useful outcomes matter to me more than making a demo look impressive for a minute.

Open source is where that habit becomes a conversation. I contribute fixes, read unfamiliar code, and try to leave a project a little easier for the next person to work with.

— Vaishnavi

01
Machine learning
Deep learning, computer vision, transformers, and practical model building.
02
AI systems
Agent tooling, inference, evaluation, and reliable developer workflows.
03
Product engineering
Full-stack apps that make everyday problems easier to handle.
04
Open source
Learning in public, reviewing code, and contributing fixes upstream.

02Selected work

The work behind
the experiments.

Explore the interfaces, then open a project to see the thinking behind it.

All repositories

forklift-safety-ai.vercel.app

KAVACH dashboard showing tracked warehouse footage

Live demo

01 / Computer vision

KAVACH

An explainable video-intelligence prototype that turns warehouse footage into tracked objects, evidence-backed safety events, and replayable incident records.

I built it around structured evidence first: YOLO perception, ByteTrack identities, temporal behaviour rules, SQLite incident history, and a grounded assistant that explains only what the evidence supports.

YOLO + ByteTrack temporal tracking
Explainable behaviour and risk scoring
SQLite evidence replay + grounded assistant

Python
OpenCV
YOLO
FastAPI

Live demo GitHub repo

github.com/Vaishnavi220506/StemPulse

STEM / PATHWAYS / RETURNBuild confidence
with evidence.

01 Learn02 Grow03 Restart

STEM pathways · evidenceView code

02 / Social impact

STEMPulse

A women-first STEM pathways platform for learning, opportunities, evidence-aware confidence, and career re-entry support.

The product maps people to a practical next step instead of treating confidence or career breaks as missing potential: profile evidence, skill freshness, opportunity matching, and a shortest-path plan.

Learn, work, fund, and re-entry pathways
Evidence-aware confidence signals
Explainable opportunity matching

React
Python
FastAPI
Career UX

GitHub repo

rescue-link-clarity-loop-ai.vercel.app

RescueLink emergency coordination home screen

Live demo

03 / Full-stack product

RescueLink

A real-time emergency coordination platform that helps people request support, offer resources, and share trusted local updates.

I designed it around a calm response workspace: role-aware access, a clear request lifecycle, location-aware maps, community confirmation, and atomic volunteer claims.

Location-aware requests and resources
Community-confirmed hazard reports
Atomic volunteer claims

React
TypeScript
Express
PostGIS

Live demo GitHub repo

clarityloop-ai.vercel.app

Live demo

04 / AI agents

ClarityLoop

An open-source workspace that stress-tests a project idea before time, money, and effort are spent building it.

Six focused agents examine requirements, evidence, resources, risks, and scope before turning the result into confidence scores, tasks, and an honest project plan.

Six-agent feasibility pipeline
Evidence and assumption tracking
Local Ollama or cloud Gemini support

Next.js
FastAPI
Gemini
Ollama

Live demo GitHub repo

raceverse.vercel.app

RaceVerse 3D racing track in the browser

Live demo

05 / ML + games

RaceVerse

A browser-based 3D racing prototype with procedural tracks, local multiplayer, AI opponents, and a trained racing policy.

Tracks are created from seeds, while the AI learns in a small simulator and runs locally in the browser. The project is a practical way to explore games, reinforcement learning, and interaction design together.

Seeded procedural tracks
Local multiplayer and AI Arena
50 unseen seeds evaluated

React
Babylon.js
Python
NumPy

Live demo GitHub repo

vaishnavi220506.github.io/ModelForge

GRAPH / REWRITE / VERIFY

.mforge→optimize→Guardian-APC+→C++ inference

99.2%

of observable injected faults caught with 128 inputs, against 75.0% for uniform random testing.

Live demo

06 / Compilers & systems

ModelForge

A C++17 tensor compiler for small neural-network models that checks every optimization before it keeps it.

It loads .mforge models, rewrites the graph, and emits standalone C++ inference code. A built-in validator, Guardian-APC+, tests each rewrite with inputs placed on the model’s ReLU boundaries and rolls back any change that alters behaviour, with a replayable counterexample.

99.2% of injected faults caught with 128 inputs
Shrink check for float16, int8, int4 and pruning
Model diff checker with terminal and browser dashboards

C++17
CMake
Compilers
ML systems

Live demo GitHub repo

And there’s more

53 repos

My own builds, plus the forks I use for open-source contributions, all live on GitHub.

All repositories

03Open source

Small fixes can makea big difference.

I learn in public by working upstream. Most of my merged work lands in computer-vision libraries like Ultralytics and Kornia, alongside fixes across AI agents, inference, ML tooling and developer workflows.

44PRs merged upstream

16Projects shipping my code

onnx/onnx Merged

ONNX

#8448 fix(shape_inference): preserve custom callback lifetime

Custom shape-inference callbacks could be freed while inference was still running. I kept guarded ownership of the callback for the full operation and added regression tests for short-lived registrations and garbage collection.

C++
ONNX
Memory safety

View PR

microsoft/PyRIT Merged

Microsoft PyRIT

#2356 FIX: support stereo audio frequency conversion

Multi-channel WAV input broke the audio frequency converter with NumPy broadcasting errors. I updated the conversion to handle stereo audio and validated it against 1,095 tests and Ruff checks.

Python
NumPy
Audio

View PR

crewAIInc/crewAI Merged

CrewAI

#6881 fix: make pre-commit hooks portable on Windows

The pre-commit hooks relied on Unix-only environment activation, so they failed on Windows. I replaced it with direct uv commands for Ruff, mypy, pip-audit and pytest, with regression coverage.

Python
AI agents
Developer tools

View PR

ultralytics/ultralytics Merged

Ultralytics YOLO

#26652 Preserve RMSprop second moments when saving checkpoints

My most active upstream project, with 14 merged PRs in the YOLO library. Recent ones keep RMSprop optimizer state in saved checkpoints, close TensorBoard writers during trainer teardown, and keep full directory names when creating ZIP archives.

Python
PyTorch
Computer vision

View PR

More merged work

See all 44 on GitHub

04The toolkit

A stack that followsthe problem.

I don’t collect tools for the sake of it. I pick the simplest set that helps an idea become clear, useful, and dependable.

Languages

The tools I reach for to think clearly and build from first principles.

Python
C++
Java
JavaScript
TypeScript
C

Web & product

A practical web toolkit for turning an idea into something people can use.

React
Next.js
Tailwind CSS
Node.js
Express
FastAPI
Vite
REST APIs
Streamlit

ML & data

Libraries and ideas for experiments that stay grounded in real constraints.

PyTorch
Hugging Face
scikit-learn
OpenCV
Ultralytics YOLO
Transformers
NumPy
Pandas
Ollama

Tools & data

The infrastructure around the code: shipping, testing, and making it dependable.

Git
GitHub Actions
Docker
PostgreSQL
PostGIS
MySQL
SQLite
SQLAlchemy
Vercel

05Achievements

A few milestonesworth keeping.

A few certificates from programs and challenges that have shaped my journey so far. Hover to see each one, and click to open its verification page.

Resume at a glance

Open full resume

EducationVIT Vellore

B.Tech in Computer Science and Engineering (SCOPE)

CGPA9.05/10

Academic foundation

InternshipIBM GenAI

Generative AI virtual internship, May – Jul 2026

Class XII · X98.2%

CBSE, Visakhapatnam · 92.8% in Class X

06Say hello

Have an idea,a question, or agood problem?

I’m always happy to talk about thoughtful software, open source, AI systems, or what you’re building next.

vaishnaviharish2006@gmail.com

I build useful things for real problems.

Curious about how things work and how to make them better.

Machine learning

AI systems

Product engineering

Open source

The work behindthe experiments.

KAVACH

STEMPulse

RescueLink

ClarityLoop

RaceVerse

ModelForge

Small fixes can makea big difference.

ONNX

Microsoft PyRIT

CrewAI

Ultralytics YOLO

More merged work

A stack that followsthe problem.

Languages

Web & product

ML & data

Tools & data

A few milestonesworth keeping.

Resume at a glance

Have an idea,a question, or agood problem?

The work behind
the experiments.