SafeChat

Here's the complete story with all seven sections — paste it as-is:

Inspiration

Every day, children are bullied in chats and games — and groomed by strangers pretending to be friends. Most of it is never reported. Kids stay silent, parents find out too late. I wanted one simple tool a child could use in the moment to ask "is this bullying?" or "is this stranger dangerous?" — and know exactly what to do next.

What it does

SafeChat — Child Safety Chat Guardian fights both dangers in one app:

Tab 1 — Check a message: paste any message → AI detects cyberbullying, names the type in plain words (age-based, gender-based, religion-based, ethnicity-based), and shows a kid-friendly "what to do now" card. Tab 2 — Check a conversation: paste a whole chat → every message gets the bullying check, plus a predator warning-sign screener rates grooming risk (low / medium / high), shows the exact trigger phrases, and gives urgent safety guidance.

Nothing is stored or sent anywhere — the child chooses what to check.

How we built it

TF-IDF + Logistic Regression (scikit-learn), trained on 40,000 labeled messages — 92% test accuracy. Instead of plain argmax, a message is flagged when the combined bullying-class probability ≥ 0.45, a threshold tuned on the held-out test set, because missing bullying is worse than a false alarm. The predator screener is a transparent pattern-based engine covering 6 categories (secrecy, meeting requests, gifts/bribes, personal-info requests, love-bombing, moving to private apps) — every flag shows its exact trigger phrase, zero black box. Built with Python and Streamlit; everything runs locally.

Challenges we ran into

An 8-hour deadline. The first model scored 92% on test data but missed softly-worded bullying — I diagnosed it with a hand-written probe set, then fixed it with the tuned 0.45 threshold, which raised both recall and accuracy. The hackathon provided no dataset, so I sourced a public labeled dataset independently. Two real training bugs (a bad CSV argument, a data-format crash) were found and fixed along the way.

Accomplishments that we're proud of

A working, deployed app — not a slide deck. 92% accuracy with per-class F1 up to 0.98. Two child-safety pillars (bullying + grooming) fused into one seamless tool. A fully transparent screener a child can understand. And honest engineering: limitations stated openly in the docs.

What we learned

End-to-end ML in one night: sourcing data, cleaning, leakage-proof train/test splits, threshold tuning, honest evaluation, and shipping. The biggest lesson: in safety-critical AI, recall beats precision — a false alarm shows a guidance card, but a missed message fails a child.

What's next for SafeChat

An ML classifier for grooming detection trained on labeled conversation data, and a privacy-preserving on-device parental dashboard — safety without surveillance.

Paste it in, add your three links, hit submit.

Built With

• joblib
• logistic-regression
• numpy
• pandas
• python
• scikit-learn
• streamlit
• tf-idf