Raj Kamal Singh

Bidirectional LSTM + attention, retrained daily on fresh data

A self-updating forecasting pipeline: a daily cron job pulls fresh OHLC prices and news headlines, scores sentiment, and retrains a Bidirectional LSTM with attention on the combined feature set — served through both a FastAPI endpoint and a Streamlit dashboard.

15% improvement in prediction accuracy over baseline models.

Two parallel CNNs with Grad-CAM interpretability

A real-time vision pipeline that detects faces via Haar Cascades, then runs two separate CNNs in parallel — one classifying gender (IMDB faces, 96% accuracy), one classifying emotion (FER-2013, 66% accuracy, matching human-level performance) — with temporal smoothing to stabilize labels and Grad-CAM visualization to interpret what each model actually learned.

96% validation accuracy on gender classification (simple CNN, IMDB faces).

Kinesis + Lambda pipeline with SNS alerting

A real-time streaming pipeline on AWS — Kinesis ingests raw data, a Lambda function scores each record with an Isolation Forest model and fires SNS alerts on anomalies, while parallel Firehose deliveries feed S3, Athena, and a QuickSight dashboard for live and historical analysis.

Processed 10K+ records/day with sub-second anomaly detection latency.

Fine-tuned FLAN-T5 at 4,000× data scale

Scaled a generative question-answering pipeline from a 20-example custom FAQ dataset to 87K+ SQuAD-style examples, fine-tuning FLAN-T5 with Hugging Face Transformers using mixed-precision training, and deploying an interactive Gradio inference app.

Scaled training data 4,000× — from 20 examples to 87K+.

ResNet-50 · home vs. restaurant vs. packaged food

Studied how a pretrained ResNet-50 classifies food images by scene context — home, restaurant, or packaged — rather than by the food itself, then compared four training strategies (baseline, augmentation, synthetic noise, and combined) to see which actually helped.

94.5% test accuracy with the plain baseline transfer-learning model — the best of the four.

Classical computer vision under adverse conditions

Built a classical edge-based lane detection pipeline (Canny + Hough transform) and stress-tested it against shadows, motion blur, and bright light, then improved robustness with adaptive thresholds, histogram equalization, and slope filtering.

Adaptive-threshold pipeline reduced false lane detections compared to the fixed-threshold baseline.

Cross-sector valuation & volatility study

A statistical analysis of S&P 500 company fundamentals — EPS, P/E, P/B, dividends, and volatility — exploring correlations between valuation metrics and how they vary across sectors.

Market Cap and EBITDA are the strongest pair in the matrix at 0.90 correlation — largely mechanical, since EBITDA drives valuation.