ml_engineer 0.99 Mohnish Kumar Shandilya

Vehicle Number Plate Detection

Detects Indian number plates with a custom-trained YOLOv8 model, reads them with plate-aware OCR and logs confirmed plates to Google Sheets. You can scan plates from a phone camera in the browser.

• YOLOv8n fine-tuned for one-class plate detection at 640 px.
• Replaced EasyOCR with PaddleOCR (PP-OCRv4). Decoding is constrained to valid Indian formats: state code, RTO digits, series and number, Bharat (BH) series and two-row plates.
• Unclear reads are flagged for review before saving. The Google Sheets log has duplicate protection and retries.
• Moved from Streamlit to a FastAPI backend and a mobile-friendly React + TypeScript app, with Docker and CI.

• Python
• YOLOv8
• PaddleOCR
• OpenCV
• FastAPI
• React
• Docker

Multi-Module Stock Prediction

AlphaStream forecasts the next day for 10 NSE large caps. It combines a forecasting model, technical rules and live FinBERT news sentiment, and a decision engine turns the three into one BUY / HOLD / SELL call.

• Compared ARIMA, XGBoost, LSTM and Transformer models, with and without news features. XGBoost + FinBERT sentiment won the walk-forward backtest and is the model served.
• Leakage-free evaluation: expanding-window chronological folds with a 5-row embargo and scalers fit per fold. Shuffling the target gives 49.9% accuracy, as it should.
• Benchmarked against random-walk, persistence, majority-direction and buy-and-hold baselines.
• FastAPI service, React + TypeScript terminal UI, MLflow tracking, Optuna tuning, Docker and GitHub Actions CI.

Honest result: no model had a reliable edge on next-day direction (49–51% accuracy), and buy-and-hold beat every model on return and Sharpe. The repo reports this rather than hiding it.

• Python
• XGBoost
• PyTorch
• FinBERT
• FastAPI
• React
• MLflow
• Docker

System Log Anomaly Detection

Flags abnormal behaviour in large-scale systems from their raw log files, so engineers hear about failures before they turn into outages. Every alert links back to the log lines behind it.

• Parses 15.9M raw log lines from the HDFS and BGL datasets into event templates with Drain, then groups them into sessions and time windows.
• Benchmarked 7 detectors (keyword rule, PCA, Isolation Forest, supervised models, a DeepLog-style LSTM and a causal Transformer) on chronological 60/20/20 splits.
• Streaming replay parses, groups and scores lines online, with detection delay measured. Adding warm-up state took HDFS precision from 0.071 to 0.968.
• FastAPI backend and React dashboard with explainable alerts. Includes Docker, MLflow, GitHub Actions and 45 automated tests.

Finding: the deep sequence models did not beat simpler methods on the test period. A decision tree reached 0.995 F1 on HDFS, and PCA was the best label-free model.

• Python
• PyTorch
• Scikit-learn
• Drain3
• FastAPI
• React
• Docker