NaukriRakshak

AI fake-job detector for freshers

Got a job offer on WhatsApp or email that feels too good to be true? Paste it here first. NaukriRakshak checks it in seconds and shows you exactly which lines are the scam — before you pay a single rupee.

Built by a final-year student who kept receiving these messages himself — "HR" on WhatsApp, refundable deposits, offer letters for a fee. Freshers and first-generation job seekers lose lakhs to fake offers every season because nobody warns them at the moment it matters: when the message is already in their hand.

The check runs entirely in your browser — a real trained model, not a chatbot. Your message is never uploaded, and there is no sign-in and no API key.

Check a job message

Paste a WhatsApp HR message, an email, an offer letter or a portal listing — anything you received.

English / Hinglish (Roman script) only for now — messages in Devanagari script can't be checked yet.

Try a sample:

Tip: the full message works best — include the salary, any fee asked, and how they contacted you.

0
Checks this session
₹0
Money saved this session
0
Messages reported
Verdict in:
0%
Scam risk
No scam signs found

Already paid or shared documents? Call 1930 and report at cybercrime.gov.in immediately, then call your bank.

The message, with danger lines marked

Why this verdict

Thanks — saved on this device only. Reports are not uploaded anywhere: the retraining loop (human-in-the-loop) needs a backend and is not built yet.

How it works

Two stages. Stage 1 is a real machine-learning model running locally in this page. Stage 2 turns the evidence into plain language anyone can act on.

STEP 1

Paste the message

A WhatsApp forward, an email, an offer letter — whatever reached you. Nothing leaves your phone or laptop.

STEP 2

A trained classifier scores it

TF-IDF (word 1–2 grams) + logistic regression, trained in scikit-learn on labelled scam vs genuine fresher postings. Its exported weights — the full 2,063-term vocabulary — run right here in your browser, scored with the exact TF-IDF maths the model was trained and validated with. The number you see is computed from the trained model, never typed in.

STEP 3

Evidence + a plain verdict

The exact suspicious phrases are highlighted and explained in simple words, with the 3 checks to do before trusting anything. In production, Gemini writes this explanation layer; this demo ships an offline explainer so the check can never fail on a network or quota error.

Tech stack

Pythonscikit-learnTF-IDF + Logistic RegressionJavaScript (in-browser inference)Gemini — explanation layer (production)

Measured accuracy — the real numbers

Held-out test on a stratified 80/20 split of the 117 labelled examples (93 train / 24 test). Reported exactly as measured — no rounding up, no borrowed benchmarks.

Held-out accuracy24 / 24 correct (1.000)
Precision (scam)1.000
Recall (scam)1.000
F1 (scam)1.000
Labelled examples117 (52 scam / 65 genuine)
Confusion matrix (held-out test)
Predicted genuinePredicted scam
Actually genuine (13)130
Actually scam (11)011

Honest caveat: 24 test examples is a small set, and they are representative examples of common patterns — real-world accuracy will be lower. That is exactly why every verdict shows its evidence and an "uncertain" band instead of pretending to be sure. Also honest: the Report button stores reports on this device only — the human-in-the-loop retraining loop needs a backend and is not built yet.