Ansh Yadav

Fake News Detection using Hybrid AI

“SHOCKING: Scientists EXPOSED the secret they don’t want you to know!!!”

Weighted verdict Likely fake Sample run, for illustration

One model can be fooled. So I made six checks argue, and weighted the verdict by what kind of text came in.

98%accuracy on fine-tuned DistilBERT

0.98F1 score

~5saverage analysis, all checks running in parallel

67Mparameters, published on Hugging Face

Try one of the six.

These are the real 8 linguistic features from my code, ported to run in your browser. They only measure style, which is why five other checks exist.

Full live demo ↗ Model ↗ Code ↗

Paste a headline

See.

PantryPilot AI

Point a phone at your vegetables. It knows what they are with no internet, because the detector runs on the phone itself.

A YOLO26s model, exported to ONNX, sits inside a Flutter app. Everything that does need a server, like receipts, recipes and expiry estimates, goes through a FastAPI backend with Gemini and MongoDB.

22vegetable classes

0network calls to detect

Code ↗

yolo26s · on device0 found

Move across the frame. A sketch of the detector, not the live model.

Guard.

MAVIS, Multi-Attack Vector Intelligence System

It learned what normal network traffic looks like. Then every alert had to explain itself.

An anomaly ensemble catches behaviour nothing has seen before. A gradient-boosted classifier names the attacks it recognises. Both share a single false-alarm budget, and every alert ships with a risk score, a reason and a MITRE ATT&CK mapping.

Watching flows

Risk0critical

Reason: One source fanning out across many ports, short flows, anomaly detectors agree
ATT&CK: T1046 Network Service Discovery
Response: Blocking source (dry run)…

0.9995ROC-AUC

0.9996recall at a 1% false-positive budget

0.982F1

2.5Mclean flows from CIC-IDS2017, a lab dataset

Code and model card ↗

Plan.

TravelGenie

Gemini returns structured JSON, a FastAPI backend checks the plan against your budget, and every stop gets a real name and a map link.

Live demo ↗ Code ↗

Waiting for a prompt ₹0 of ₹40,000

Sample plan, for illustration

Now you teach one.

This is a real neural network with real backpropagation, training live in your browser. Click to give it examples. It has to figure out the rest.

Click anywhere to add a point

Paint with

Or load a dataset

Learning rate 0.03

Lossn/a

Accuracyn/a

Steps0

2 → 12 → 12 → 1, tanh, plain SGD with momentum. No libraries.

Teaching people too

Celora is the AI-assisted learning platform I’m building now for students and educators, on the MERN stack. Authentication and the core learning flows are in. It’s still in progress.

Also built.

Model card.

MAVIS shipped with one. It seemed fair to write my own.

ansh-yadavv3 · still training

Model details

Third-year B.Tech Computer Science student specialising in AI and ML at Vellore Institute of Technology, Chennai. Expected to finish training in 2028.

Intended use

AI and ML engineering roles and internships, especially where a model has to survive contact with real users.

Out of scope

Fixing your printer.

Frameworks

PyTorch, scikit-learn, Hugging Face Transformers. Python, C++, Java, JavaScript, SQL. React, Node.js, Express, Flask, FastAPI, Firebase, MongoDB.

Evaluation

CGPA 8.22 / 10.00. Selected participant, Smart India Hackathon, college level, 2025.

Other training data

Media and outreach for the Entrepreneurship Cell, VIT Chennai: Pitcher’s Showdown, Stock Market Simulation, IPL Battle. Certificates in generative AI (Microsoft and LinkedIn), Azure AI, intro to AI and cloud computing.

Known limitations

Classifies poorly under adversarial clicks on his name.
Up, up, down, down causes severe regularisation.
Output gets strange at high temperature. See the footer.
The small yellow one should not be shaken.
Responds to the word “hire”, typed anywhere.
Developers should open the console.

Say hi.

LinkedIn ↗ GitHub ↗ Resume ↗