Sheikh
Khairul Momin
Mohammad Tahmid

CriticLens: Critic vs Audience Sentiment Divergence Engine

End-to-end NLP platform quantifying critic-vs-audience sentiment divergence across 19,836 films and 1M+ reviews via a custom divergence scoring algorithm built on Rotten Tomatoes tomatometer and audience score deltas. Fine-tuned DistilBERT on 50K IMDB reviews for binary sentiment classification, with a TF-IDF + SGD fallback for low-latency inference — both models serving live predictions cached in MySQL. Automated multi-source ingestion pipelines pull from TMDB, OMDb, The Guardian, and NYT Article Search APIs via APScheduler, with a 69-month historical backfill yielding 2,600+ critic reviews. Schema migrated from MySQL to TiDB Cloud Serverless with SSL-secured connections and smart startup sequencing. Fully containerised with Docker; FastAPI backend serves a vanilla JS frontend with film search, divergence leaderboard, sentiment timeline, and aspect-level critic breakdown.

Explainable AI Hiring Intelligence Platform

Modular 12-app Django REST API with JWT auth and role-based permissions, generating 384-dim SBERT embeddings to cosine-rank 3,000+ candidates per job posting. SHAP and LIME explanations are fused 60/40 into a unified explainability report per match, with disparate impact ratio and 4/5 rule computed across gender, ethnicity, and age groups. An async Celery pipeline orchestrates CV parsing, embedding generation, and batch matching, while a trainable GradientBoosting scorer ships with zero-downtime fallback to fixed weights.

TrialGuard: Clinical Trial Patient Dropout Prediction Platform

Django SaaS platform predicting clinical trial dropout up to 60 days in advance. XGBoost classifier with per-visit SHAP waterfall plots identifies at-risk patients at the individual level. Cox Proportional Hazards model delivers hazard ratios and 30/60/90-day retention probabilities, with stratified Kaplan-Meier curves segmented by risk tier (Low to Critical) via Lifelines. Cohort forecasts with 95% confidence intervals served through a DRF REST API; branded ReportLab PDFs generated per patient including SHAP feature drivers, survival curves, and clinical action logs.

Contract Intelligence & Power Imbalance Platform

Fine-tuned LegalBERT and Legal-RoBERTa-large ensemble classifying every clause across a 100-type unified taxonomy (CUAD, LEDGAR, MAUD - 132k clauses, macro F1 0.606). Anomaly detection fuses Isolation Forest with a 1024-dim autoencoder; bilateral power imbalance is scored −100 to +100 via sentiment, modal verb, and obligation NLP. Every prediction is explained at token level with SHAP, served through a FastAPI REST API and dark-theme dashboard with downloadable PDF reports.

AI Revenue Leakage Detection Platform

Production contract intelligence platform with XGBoost/Logistic Regression ensemble (70/30) scoring invoice leakage probability, Isolation Forest anomaly detection per record, and Prophet forecasting with 90% CI bands over a 24-month revenue horizon. SHAP attribution pipeline exposes top risk drivers per invoice; a 7-rule leakage engine generates 13k+ alerts covering missing payments, underbilling, and duplicates. Predictions served via DRF API with a Chart.js dashboard and per-invoice drill-down modals.

CostRadar: AWS Spend Anomaly Detection and Rightsizing Platform

Production AWS cost intelligence platform ingesting 90-day per-service spend via Cost Explorer API into DynamoDB. Dual anomaly detection using Isolation Forest (contamination=0.05) and Z-score with consensus scoring and plain-English explanations. Prophet forecasting with 30-day horizon, 95% CI, MAPE validation, and budget breach day prediction. EC2 rightsizing via 14-day CloudWatch CPU/memory analysis. EventBridge-scheduled daily Lambda pipeline with SNS alerts, containerised via Docker and deployed to Hugging Face Spaces.

AI Portfolio Stress Testing Platform

Production risk platform with GMM regime classification, Ledoit-Wolf shrinkage, and XGBoost/ElasticNet trained on 50+ macro features, delivering Sharpe 4.42, VaR -12.07%, and CVaR -17.48%. Stress-tested 20+ historical crises with SHAP attribution and plain-English narratives, plus reverse stress-testing for threshold-based shock back-solving. FastAPI dashboard refreshes 20 live instruments hourly.

PolicyGate: AI Agent Action Firewall

Deployed a framework-agnostic LLM agent firewall with a YAML policy DSL (6 operators, hot-reload on file mtime), dual-layer safety enforcement via system prompt and mechanical gate, and per-response confidence scoring (High/Medium/Low) via constrained secondary LLM call. Supports LangChain, CrewAI, and AutoGen via POST /gate. All gate decisions logged to SQLite with WAL mode and threading.Lock() for concurrent write safety.

AI Lead Quality Scoring & Spam Detection

Fine-tuned BERT (99.12%) and RoBERTa (99.41%) for spam detection, paired with XGBoost scoring at R²=0.9996 across 4,137 engineered features from UK Companies House, firmographics, and DNS/SMTP signals. Deployed via FastAPI with live SHAP explanations, a Tailwind SPA, and an automated nightly retraining pipeline.

Done in collaboration with Prince Kumar Nath.

Spam Lens: Explainable Spam Classifier with Multi-Method XAI

Dual-stream BiLSTM with a custom Attention layer blends token embeddings with 5 linguistic signals (URL, phone, currency, urgency ratio, length norm); Attention, LIME, and SHAP are fused via per-method normalisation into a single word-importance ranking, plain-English verdict, and live per-token colour heatmap. Deployed on FastAPI with concurrent XAI thread pools, SQLite history, and a live analytics dashboard — retraining triggers automatically on error rate, staleness, or volume thresholds, with accuracy-gated rollback.