← Home

Projects

Production-focused ML projects — end-to-end system design from model development through deployment and monitoring.

---

Time Series Forecasting at Scale

Stack: PyTorch, Hydra, MLflow, Optuna, FastAPI, Docker, Streamlit Result: 7.1% MAPE on Walmart M5 (28-day horizon, walk-forward validation, no data leakage) Links: Repo | Demo (~30s cold start)

Production-grade forecasting portfolio with reusable ML components across retail sales, rideshare demand, and inventory planning.

Challenge

• Forecasting solutions need to generalize across diverse business domains without rewriting pipelines
• Fair model comparison requires controlled experiments: same data splits, same metrics, same preprocessing
• Scaling to 80K+ time series samples while keeping experiments reproducible

Architecture Decisions

• Hydra for config management — every experiment is fully reproducible from a single YAML file. No hard-coded parameters.
• MLflow for tracking — all runs logged with metrics, parameters, and artifacts. Trivial to compare statistical vs. deep learning approaches across domains.
• Optuna for HPO — Bayesian optimization over grid search. Pruning cuts unpromising trials early, saving compute on the 10+ architecture sweep.
• FastAPI for serving — prediction endpoints with confidence intervals, not just point forecasts. Dockerized for consistent deployment.
• Streamlit for stakeholder demos — interactive dashboard lets non-ML users explore forecasts and compare models visually.

Model Zoo

• Statistical: LightGBM, XGBoost, CatBoost, Prophet, ARIMA, SARIMA
• Deep Learning: LSTM, TCN, Transformer
• Domains: Walmart M5 retail, Ola rideshare, inventory planning, Transportation TSI

Results

• 7.1% MAPE on Walmart retail (28-day forecast horizon, walk-forward validation)
• 80K+ time series samples across 4 real-world projects
• Production pipeline: data ingestion → training → serving (Dockerized)
• 10+ architectures benchmarked with tracked experiments

---

Movie Recommendation System

Stack: PyTorch, Hydra, MLflow, Optuna, Gradio, Streamlit Result: RMSE ~0.323 on MovieLens (610 users, 9,742 movies, 100K+ ratings) Links: Repo | Demo (~30s cold start)

End-to-end recommendation system — full ML lifecycle from training through experiment tracking to production deployment with global API access.

Challenge

• Building multiple recommendation approaches that can be fairly compared under identical conditions
• Making ML demos accessible to non-technical users while maintaining engineering rigor
• Deploying a live system with low-latency inference

Architecture Decisions

• 4 model architectures for systematic comparison — deep collaborative filtering (neural matrix factorization), hybrid (collaborative + content signals), traditional collaborative filtering, and content-based. Each shares the same training/eval harness.
• Hydra + MLflow + Optuna — same reproducibility stack as the forecasting project. Consistent tooling across projects reduces cognitive overhead.
• Dual deployment — Gradio for quick interactive demos, Streamlit for richer exploration. Both served from the same model artifacts.
• Hugging Face Spaces — zero-ops deployment with global CDN.

Results

• RMSE ~0.323 on MovieLens test set
• Live deployment on Hugging Face Spaces
• Full ML lifecycle: data pipeline → training → HPO → serving
• Multiple interfaces: Gradio, Streamlit, REST API

---

Floorplan Wall Segmentation (2D to 3D)

Stack: PyTorch, DeepLabV3, OpenCV, Hydra, ONNX Result: 99.4% pixel accuracy — wall-class F1 significantly above baseline despite 96/4 class split, 50ms inference Links: Repo

Computer vision for semantic segmentation of architectural floorplans — identifying walls to enable automated 2D-to-3D conversion.

Challenge

• Extreme class imbalance in RPLAN dataset: 96.2% background, 3.7% walls, 0.01% rooms
• Processing 80K+ architectural floorplan images (64K train, 16K val)
• Accurate wall boundary detection needed for downstream 3D reconstruction

Architecture Decisions

• DeepLabV3 + ResNet50 — atrous convolution captures multi-scale context, critical for walls that vary from thin lines to thick structural elements. Transfer learning from ImageNet despite the domain gap.
• Class weighting — standard cross-entropy would ignore the 3.7% wall class. Inverse-frequency weighting forces the model to learn wall boundaries.
• OpenCV postprocessing — morphological closing/opening + connected component analysis cleans up noisy predictions. Cheaper and more interpretable than a larger model.
• Mixed precision (AMP) — halves GPU memory, enabling larger batch sizes on the 80K dataset.
• ONNX export — decouples inference from PyTorch. Enables deployment on edge devices or non-Python backends.

Results

• 99.4% pixel accuracy — note: with 96% background, pixel accuracy alone is misleading. The model’s value is in wall-class performance (F1/Dice significantly above a naive baseline)
• 80K+ images: 64K training, 16K validation
• 50ms inference — fast enough for interactive use
• ONNX export for cross-platform deployment

---

More Projects

GitHub profile for additional projects and contributions.