SMNA.
Dhaka --:-- GMT+6

Fig. 114,000 pts
A forward pass moving through five layers.

S. M. Nihal Ahmed. Software Engineering student majoring in Data Science at Daffodil International University, Dhaka. I research explainable AI and AI for cybersecurity, then ship the models as real services.

Drag the seam, or pick a side
Side A · Why it works

I want to know why a model decides what it decides. My first paper tests three gradient-boosting models for phishing detection on four datasets, then checks whether their SHAP explanations hold up under attack.

Journal of Cyber SecurityVol. 8 No. 1pp. 667–7032026

Explainable Gradient Boosting for Transparent Phishing URL Detection: A Cross-Dataset, Statistically Validated SHAP Analysis of XGBoost, LightGBM, and CatBoost

S. M. Nihal Ahmed, Afrim Hossen Khan, Md. Mazbaur Rashid

First author Corresponding author Read paper ↗ Scholar ↗
99.11%XGBoost test accuracy, dataset A
0.999ROC-AUC
4datasets, 2015 to 2026
Key finding

A broader SHAP evidence base did not mean robustness. XGBoost was the easiest model to fool with SHAP-guided attacks.

XGBoost
50.0%
CatBoost
33.9%
LightGBM
32.4%
Mean evasion rate, phishing URLs misread as legitimate. Table 16, four datasets.
Mean |SHAP| · dataset A · top 10
1PctExtHyperlinks1.410
2PctNullSelfRedirectHyperlinks0.970
3PctExtNullSelfRedirectHyperlinksRT0.945
4PctExtResourceUrls0.857
5PathLevel0.723
6NumDash0.688
7FrequentDomainNameMismatch0.687
8InsecureForms0.668
9NumNumericChars0.629
10NumDots0.592
00.511.5
From Table 11 of the paper. All three models rank the same ten features, but weigh them differently. Switch models to watch the order change.
Research interests
  • Computer vision
  • Explainable AI
  • AI for cybersecurity
  • Neuro-inspired control
  • Medical imaging
  • Model compression
Side B · How it ships

A model is only useful once someone can call it. I package what I train into small, fast services with real APIs, then shrink them until they run anywhere. Four of them are live web demos right now.

4live web demos
4models on Hugging Face
3ONNX exports for CPU
Live right now PromptGuardpromptguard-xofd.onrender.com↗ Face Authenticity Scannerdeepfake-face-scanner.onrender.com↗ GameSensegamesense-h456.onrender.com↗ Bangla SMS Detectorsms-spam-detector.onrender.com↗

Free hosting, so a demo can take about 30 seconds to wake up.

PyTorchTensorFlowScikit-learnHugging FaceONNXFastAPIFlaskDockerGit
How a model ships
  1. TrainPyTorch, TensorFlow or scikit-learn, usually with transfer learning
  2. ExplainSHAP or LIME to check what the model actually relies on
  3. Shrinkexport to ONNX, then quantize, prune or distill
  4. ServeFastAPI or Flask in Docker, deployed on Render

Models thatleave the notebook.

Seven projects, from robot arms wired like a fruit fly's brain to a Bangla SMS filter. Each one is trained, tested and shipped as something you can open. Free-tier demos can take about 30 seconds to wake up.

  • 2026Explainable phishing detection First-author journal paper. Three gradient-boosting models, four datasets, McNemar tests, and a SHAP-guided attack that tests whether explanations predict robustness.XGBoost · LightGBM · CatBoost · SHAP60:20:20 leakage-free protocol · Journal of Cyber SecurityXGBoost · LightGBM · CatBoost · SHAP99.11% acc · 0.999 AUC PaperPaperScholar
  • 2026Fly-brain robots Robot arms run by a recurrent network wired like a fruit fly's brain, where only the connection strengths are learned. It picks ripe tomatoes and leaves the unripe ones.FlyWire connectome · MuJoCo · PPO + DAggerFlyWire connectome v783 · recurrent PPO + DAggerPython · PyTorch · Stable-Baselines3 · MuJoCo98% ripe picks · 0 of 400 unripe ResearchProjectCode
  • 2026PromptGuard Flags prompt-injection and jailbreak attempts before they reach a language model.mBERT · ONNX Runtime · FlaskFine-tuned mBERT · 567-prompt test splitONNX Runtime · Flask · Docker · live on Render88.4% acc · 0.962 AUC SecurityDemoCodeModel
  • 2026Face Authenticity Scanner Tells real faces from AI-generated ones by fusing two CNN backbones into one classifier.EfficientNet-B1 + RegNetY-800MF · ONNXDual-backbone feature fusion · 21,316-face test splitONNX · Flask · weights from Hugging Face · live on Render95.4% acc · 0.990 AUC VisionDemoCodeModel
  • 2026GameSense Names the game from a single screenshot, across 10 titles from Minecraft to Genshin Impact.EfficientNet-B0 · FastAPITransfer learning · 10 classes · 1,500 test framesFastAPI · live on Render99.93% acc · 1,500 test frames VisionDemoModel
  • 2026Bangla SMS Detector Sorts Bangla text messages into normal, promotional and smishing.TF-IDF · LinearSVC · FastAPITF-IDF n-grams · 5-fold CV · 555-message test setFastAPI · scikit-learn · live on Render98.2% test accuracy SecurityDemoCode
  • EarlierRice Disease Bot A Telegram bot that spots four rice leaf diseases from a photo and explains treatment in Bangla.ResNet50 · ONNX · Telegram4 classes · 5,232 images, class-balancedONNX · Telegram bot · replies in Bangla98.47% best val accuracy VisionPreviewCodeModel
  • EarlierMango Leaf Disease Classifies eight mango leaf diseases, and uses LIME to show which parts of the leaf drove each call.ResNet50 · LIME · 4,000 images8 classes · 4,000 images · LIME explanationsJupyter notebook · trained model in GitHub releases VisionCode

Education

3.70 / 4.00 CGPA

B.Sc. in Software Engineering, major in Data Science. Daffodil International University, Dhaka. 2023 to 2027 (expected).

Bangla (native) · English (professional working proficiency)

Publication

JCS 2026

Explainable gradient boosting for transparent phishing URL detection. Journal of Cyber Security, Vol. 8 No. 1, pp. 667–703. First and corresponding author.

DOI · Google Scholar · ResearchGate

On Hugging Face

4 models

Leadership

3MT 2025

Organizer of the DIU SWE Three Minute Thesis competition, February 2025. Recognized for organizing the university's 3MT event, where researchers explain their work in three minutes.

Certifications · tap to verify

Got a modelthat needs tosee, reador defend?

[email protected]
GitHub ↗ LinkedIn ↗ ORCID ↗ Google Scholar ↗ ResearchGate ↗ Hugging Face ↗ Paper ↗