ML / AI Engineer · Computer Vision · Published Researcher · Original Dataset Author
I build ML/DL systems from scratch — hand-coded backpropagation, diffusion samplers, camera calibration via DLT+SVD, and biometric evaluation pipelines. ~5 years of experience across computer vision, generative modeling, NLP, and production software engineering.
Languages: Python SQL MATLAB R C JavaScript
DL / ML: PyTorch TensorFlow scikit-learn · CNNs · LSTMs · Transformers · Diffusion (DDIM) · GANs · Autoencoders
Computer Vision: OpenCV · YOLO · Face Recognition · Biometrics (PAD) · Medical Imaging
Data & Infra: pandas NumPy Docker Git Linux Supabase/Postgres REST APIs CUDA HuggingFace MLflow
Cloud: AWS Solutions Architect (Associate) · AWS Cloud Practitioner · AWS AI Practitioner
These are the highlights — all 23 repos are listed below
| Project | What I Built | Key Result | |
|---|---|---|---|
| 🎨 | DDIM Inpainting | From-scratch UNet + DDIM diffusion with v-prediction, self-conditioning, EMA, CoordConv | PSNR 14.41 on MNIST |
| 🔐 | Face PAD (Anti-Spoofing) | End-to-end video-level face attack detection. Subject-disjoint protocol, EER/ACER thresholds | ACER 0.029, Acc 98.1% |
| 🍇 | Blackberry-Lime Classifier | Original 6K-image dataset (published on Kaggle) + dual ResNet-18 classifier → Live App | 100% / 99% accuracy |
| 🧠 | Transformer from Scratch | Encoder + Decoder with sinusoidal positional encoding, multi-head attention on Yelp Polarity | 79% test, ROC-AUC |
| 📈 | LSTM Stock + VaR | Next-day forecasting with Value-at-Risk quantification on Dropbox stock | MAPE 1.51%, R² 0.908 |
| 🧮 | From-Scratch MLP | Hand-coded forward pass + backpropagation + L2 reg, no autograd. MNIST + CelebA faces | 95% / 85% test |
| 📹 | YOLOv3 Traffic Detection | Real-time detection on traffic video with NMS, custom IoU, FPS profiling | Real-time on 3 videos |
| 📷 | Camera Calibration | DLT + SVD camera calibration from scratch — intrinsic & extrinsic decomposition | Pure NumPy math |
"Cardiovascular Disease Prediction using ML" — IJSREM, Vol. 07, Issue 08, Aug 2023
Co-authored · Cleveland Heart Disease DB · LogReg vs Random Forest · Flask web app
📎 View Paper
🧠 Deep Learning & Generative Models (click to expand)
- ddim-mnist-inpainting — DDIM diffusion, UNet, v-prediction, self-conditioning
- transformer-from-scratch-seq — Transformer encoder+decoder on Yelp Polarity
- lstm-sentiment-analysis — LSTM vs GloVe+BiGRU on IMDB 50K (81.8%)
- anomaly-detection-autoencoders — Dense AE on NAB CloudWatch data
- lstm-stock-forecasting-with-var — LSTM + Value-at-Risk on stock data
- covid-mlp — MLP on CDC COVID data (87.58%)
👁️ Computer Vision & Biometrics (click to expand)
- face-pad-msu-mfsd — Video-level face anti-spoofing (PAD)
- faces-can-lie-identity-analysis — ArcFace embeddings + clustering
- traffic-object-detection-yolov3 — Real-time YOLOv3 detection
- camera-geometry-calibration — DLT camera calibration from scratch
- rgb-channel-alignment — ORB + RANSAC homography alignment
- blackberry-lime-visual-classifier — Original dataset + deployed app
- octmnist-retinal-disease-cnn — Medical imaging CNN (91.3%)
- vgg16-vs-resnet18-64x64 — From-scratch VGG vs ResNet
- cnn-multiclass-28x28 — 36-class CNN, 3-optimizer comparison
📊 Classical ML (From Scratch — No sklearn) (click to expand)
- mnist-logreg-svm — Hand-coded LogReg + Softmax + SVM sweep (98.32%)
- lda-qda-ridge-regression — LDA/QDA/Ridge all from scratch
- handwritten-digits-and-faces-mlp — MLP with manual backprop
- kmeans-farthest-first-init — Farthest-first K-Means (3.4× speedup)
- adult-census-income-prediction — NB/DT/RF/XGBoost pipeline
🔬 Research & Numerical Methods (click to expand)
- nanoglass-md-preprocessor — MATLAB + LAMMPS for Cu-Zr nanoglass MD (B.Tech thesis)
- predictive-health-assessment — K-Means health risk stratification (R, 100K records)
- Mini-Projects — Gauss-Newton, Monte Carlo/LHS, Newton-Raphson, ARIMA, time-series
Master of Science in Engineering Science (Data Science) — University at Buffalo (SUNY), 2024–2025
Bachelor of Tech — IIT Bhubaneswar, 2019–2023
Building models that see, think, and act intelligently.