Applied Machine Learning Projects

A monorepo of nine end-to-end machine learning projects spanning computer vision, large language models, graph learning, time-series forecasting, and production inference serving. Each project lives in its own subdirectory with independent dependencies, tests, documented results, and a full README.

Projects at a Glance

Project	Area	Key Technologies	Standout
ml-satellite-image-classifier	Computer Vision	PyTorch, Keras/TF, FastAPI, Docker	99.83% accuracy; FastAPI server serving all four models
ml-llm-alignment-fine-tuning	LLM Alignment	PyTorch, TRL, HuggingFace, LoRA	Full SFT → RM → PPO RLHF → DPO pipeline, all trained locally
ml-tiny-llm-gpt	Language Modeling	PyTorch	GPT decoder-only Transformer built from scratch
ml-gcp-vertex-rag-chatbot	RAG / GenAI	LangChain, Vertex AI, Chroma, Cloud Run	Document Q&A app deployed to GCP Cloud Run
ml-movie-recommender	Graph ML	PyTorch Geometric, igraph	Heterogeneous GNN over IMDb graphs; top-N recommendation on MovieLens
ml-social-network-predictor	Graph ML	igraph, PyTorch	DeepWalk embeddings reach 0.986 ROC-AUC on 4,039-node Facebook graph
ml-wearable-motion-classifier	Classical ML	scikit-learn, NumPy, SciPy	IMU → trajectory → ensemble classifier for clinical rehabilitation
ml-recyclable-material-classifier-vgg16	Computer Vision	Keras/TF, VGG16	Transfer learning; caught and fixed training-on-test bug from source notebook
ml-boston-climate-modeler	Time-Series	TensorFlow, Python	Pure-TF LSTM + Transformer from scratch (no Keras); 7-day multi-step forecasting; 56 unit tests

---

Project Details

ml-satellite-image-classifier

Binary classification of 64×64 satellite image tiles as agricultural vs. non-agricultural land.

• Models: Six-block CNN (32→1024 channels, 5×5 kernels) and CNN–Vision Transformer hybrid (CNN feature map tokenized and fed through multi-head self-attention blocks), each implemented independently in both Keras/TensorFlow and PyTorch — four models total, trained and evaluated in parallel across frameworks.
• Results: PyTorch CNN 99.83%, Keras CNN 99.33%, PyTorch CNN-ViT 99.67%, Keras CNN-ViT 99.42% — all on a 1,200-image held-out split never seen during training.
• Inference server: FastAPI app (serve/) loads all four models at startup and exposes /health, /models, and POST /predict?model= endpoints. Model backend is selectable per request. Deployed with Docker Compose; model weights mounted read-only at runtime to keep the image small.
• Notable: Caught and fixed a data-leakage bug in the original evaluation methodology that scored models against the full training set rather than a held-out split.

Stack: Python · PyTorch · Keras/TensorFlow · FastAPI · Uvicorn · Docker Compose

---

ml-llm-alignment-fine-tuning

Four LLM alignment techniques implemented end-to-end, each with its own training objective and evaluation metric — all locally runnable on a laptop CPU.

• Supervised fine-tuning (SFT): LoRA-adapts facebook/opt-350m on CodeAlpaca-20k with TRL's SFTTrainer; evaluated with SacreBLEU before and after fine-tuning.
• Reward modeling: GPT-2 + LoRA trained on chosen/rejected response pairs with RewardTrainer; reaches 0.96 pairwise ranking accuracy on a held-out preference set.
• PPO RLHF: gpt2-imdb steered toward positive and negative sentiment with PPOTrainer against a sentiment-classifier reward; mean reward improves from 0.24 → 1.27 (positive policy) and −0.32 → 0.56 (negative policy).
• DPO: GPT-2 + LoRA fine-tuned directly on preference pairs with DPOTrainer (no separate reward model or RL loop); reaches 0.70 reward accuracy on held-out pairs.
• Every script replaces commented-out or pre-downloaded training from the source notebooks with real, locally-runnable training — every number above comes from an actual training run.

Stack: Python · PyTorch · HuggingFace Transformers · TRL · LoRA (PEFT)

---

ml-tiny-llm-gpt

A from-scratch GPT-style language model covering the complete pipeline from raw text to generated output.

• Architecture: Decoder-only Transformer with configurable depth, heads, and embedding dimension; causal self-attention, learned positional embeddings, and layer normalization — no HuggingFace model code in the loop.
• Pipeline: Custom BPE tokenizer training → dataset preprocessing and sequence packing → training loop with gradient clipping and validation checkpointing → top-k / top-p text generation → perplexity evaluation → throughput and memory benchmarking.
• Intentionally sized to run on consumer hardware while demonstrating every component of a modern LLM training stack.

Stack: Python · PyTorch

---

ml-gcp-vertex-rag-chatbot

A deployed RAG document Q&A app: upload a PDF, TXT, Markdown, CSV, or DOCX file and ask questions grounded in its content.

• RAG pipeline: LangChain document loaders → RecursiveCharacterTextSplitter → Vertex AI text-embedding-004 embeddings → Chroma vector store → RetrievalQA chain → Gemini 2.5 Flash answer with source grounding.
• Deployment: Containerized with Docker and deployed to GCP Cloud Run with scale-to-zero cost controls; credentials handled via Application Default Credentials for local development.
• Interface: Gradio web UI; standalone annotated scripts for each RAG concept (loading, splitting, embedding, retrieval) as reference implementations.

Stack: Python · LangChain · Google Vertex AI (Gemini + text-embedding-004) · Chroma · Gradio · Docker · Cloud Run

---

ml-movie-recommender

Graph feature engineering pipeline extended with a heterogeneous GNN, evaluated on two separate tasks.

• Feature engineering (igraph): Actor/movie networks built from IMDb data; actors ranked by PageRank, movie communities detected with Fast Greedy Newman, Jaccard movie-movie similarity computed — all kept exactly as in the original coursework pipeline.
• Heterogeneous GNN (PyTorch Geometric): HeteroConv encoder with GraphConv for weighted relations and SAGEConv for unweighted bipartite edges; graph-derived features (PageRank, community ID, Jaccard similarity) used as node features.
• IMDb track: Movie rating prediction benchmarked against neighborhood averaging, linear regression, and bipartite graph averaging baselines.
• MovieLens track: Genuine personalized top-N recommendation on ml-latest-small (943 users, ~9,700 movies), evaluated with Precision/Recall/NDCG@{5,10,20} against the full catalog — not sampled negatives.

Stack: Python · PyTorch · PyTorch Geometric · igraph · scikit-learn

---

ml-social-network-predictor

Structural analysis of large social graphs extended into a link-prediction task comparing hand-engineered heuristics against learned node embeddings.

• Graph analysis (igraph): Degree distribution, ego-network extraction, Fast-Greedy / Edge-Betweenness / Infomap / Walktrap community detection, embeddedness and dispersion scoring on 4,039-node Facebook and Google+ graphs.
• Node embeddings: Hand-rolled DeepWalk-style skip-gram model (PyTorch) trained on random walks from the training graph only — test edges withheld before any graph operation.
• Link prediction results (full Facebook graph, 1,000 held-out test edges): heuristics-only 0.974 ROC-AUC, embeddings-only 0.986 ROC-AUC, combined 0.953 ROC-AUC (threshold sensitivity; see case study writeup).
• Proper holdout methodology documented: test edges excluded before all graph analysis and embedding training to prevent leakage.

Stack: Python · PyTorch · igraph · scikit-learn

---

ml-wearable-motion-classifier

A signal-processing and ML pipeline classifying upper-body movements from wrist-worn IMU data in a clinical rehabilitation context (Wolf Motor Function Test).

• Preprocessing: Sensor-frame alignment, gravity subtraction, zero-velocity updates (ZUPT), and wrist trajectory reconstruction from raw accelerometer/gyroscope/quaternion captures of a wrist-worn MPU-9150.
• Features: Vertical power, azimuth rotation, peak counts, path length, variance, and acceleration statistics extracted from reconstructed trajectories.
• Models: Deterministic rule-based baseline + four scikit-learn classifiers (SVM, random forest, histogram gradient boosting, soft-voting ensemble). Data augmentation pipeline with source-trial-aware grouped cross-validation to prevent leakage from augmented variants back into the test fold.
• Results: ~74% accuracy (grouped CV, 15 of 17 classes, real + augmented data) — see README for an honest discussion of dataset limitations.
• Testing: 6-module test suite covering the CLI, sensor parser, preprocessing, quaternion math, feature extraction, and the rule-based classifier. Not approved for clinical use.

Stack: Python · scikit-learn · NumPy · SciPy · pandas

---

ml-recyclable-material-classifier-vgg16

Binary classification of organic vs. recyclable material images using VGG16 transfer learning, with two configurations compared head-to-head.

• Configurations: Frozen feature-extraction head vs. fine-tuning from block5_conv3 onward; both trained on 800 images with augmentation, evaluated on a 200-image held-out test set.
• Results: Both reach 86.5% test accuracy; fine-tuning improves ROC-AUC (0.9534 vs. 0.9332) and produces more balanced precision/recall across classes.
• Notable: Caught and fixed a data-handling bug in the source notebook — a mislabeled generator meant every original training run was training on test-directory images, not training-directory images.

Stack: Python · Keras/TensorFlow · VGG16

---

ml-boston-climate-modeler

Daily weather forecasting for Reading, MA (Boston suburb) from NOAA station data — two complete pipelines in one repo: a stdlib-only Ridge baseline and a pure-TensorFlow deep learning stack.

• Ridge baseline (v0.1): NOAA CSV parsing → missing-value handling → seasonal, lag, and rolling-window feature engineering → Ridge regression from scratch (Gauss-Jordan solver, no NumPy) → model serialization to JSON. Temperature R² = 0.747 on 2016 test set.
• LSTM forecaster (v0.2): 2-layer stacked LSTMCell unrolled with tf.unstack; all gate weights as raw tf.Variable; trained with tf.GradientTape and a hand-built Adam optimiser. Predicts PRCP, SNOW, and TOBS jointly 7 days ahead.
• Transformer forecaster (v0.2): Pre-norm encoder with sinusoidal positional encoding, MultiHeadAttention (Q/K/V projections as tf.Variable, scaled dot-product via tf.linalg.matmul), GELU feed-forward blocks, and mean pooling. No tf.keras API used anywhere.
• Training infrastructure: tf.GradientTape mini-batch loop, tf.clip_by_global_norm gradient clipping, temporal validation split, early stopping. Trained on 58 years of data (1960–2017, 20 974 windows); the Transformer reaches R² = 0.693 on 7-day-ahead temperature over a 2018–2019 test set, outperforming the LSTM (R² = 0.639). PRCP/SNOW R² near zero is expected given event sparsity.
• Notebook: notebooks/climate_exploration.ipynb covers EDA, training loss curves, attention heatmap visualisation for each encoder block, and a four-model comparison.
• Testing: 56 unit tests across data cleaning, feature engineering, sequence windowing, scaler round-trips, TF primitive shapes, and weight serialisation.

Stack: Python · TensorFlow 2.14+ (tf.Module / tf.Variable / tf.GradientTape) · NumPy · Jupyter · Matplotlib

---

References

Papers and resources that directly informed the techniques used across these projects.

Transformers and attention

• Vaswani, A., et al. "Attention Is All You Need." NeurIPS, 2017. arxiv.org/abs/1706.03762
• Dosovitskiy, A., et al. "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale." ICLR, 2021. arxiv.org/abs/2010.11929 (ml-satellite-image-classifier)

Recurrent networks and time-series forecasting

• Hochreiter, S., and Schmidhuber, J. "Long Short-Term Memory." Neural Computation, 9(8):1735–1780, 1997. doi.org/10.1162/neco.1997.9.8.1735 (ml-boston-climate-modeler)
• Jozefowicz, R., et al. "An Empirical Exploration of Recurrent Network Architectures." ICML, 2015. proceedings.mlr.press/v37/jozefowicz15.html (ml-boston-climate-modeler — forget-gate bias initialisation)
• Kingma, D.P., and Ba, J. "Adam: A Method for Stochastic Optimization." ICLR, 2015. arxiv.org/abs/1412.6980 (ml-boston-climate-modeler)
• Ba, J.L., et al. "Layer Normalization." 2016. arxiv.org/abs/1607.06450 (ml-boston-climate-modeler)

Language modeling

• Radford, A., et al. "Language Models are Unsupervised Multitask Learners." OpenAI, 2019. openai.com/research/language-unsupervised (ml-tiny-llm-gpt)
• Eldan, R., and Li, Y. "TinyStories: How Small Can Language Models Be and Still Speak Coherent English?" 2023. arxiv.org/abs/2305.07759 (ml-tiny-llm-gpt)
• Karpathy, A. nanoGPT. github.com/karpathy/nanoGPT (ml-tiny-llm-gpt)

LLM alignment and fine-tuning

• Ouyang, L., et al. "Training Language Models to Follow Instructions with Human Feedback." NeurIPS, 2022. arxiv.org/abs/2203.02155 (ml-llm-alignment-fine-tuning)
• Rafailov, R., et al. "Direct Preference Optimization: Your Language Model is Secretly a Reward Model." NeurIPS, 2023. arxiv.org/abs/2305.18290 (ml-llm-alignment-fine-tuning)
• Hu, E., et al. "LoRA: Low-Rank Adaptation of Large Language Models." ICLR, 2022. arxiv.org/abs/2106.09685 (ml-llm-alignment-fine-tuning)
• Schulman, J., et al. "Proximal Policy Optimization Algorithms." 2017. arxiv.org/abs/1707.06347 (ml-llm-alignment-fine-tuning)

Retrieval-augmented generation

• Lewis, P., et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." NeurIPS, 2020. arxiv.org/abs/2005.11401 (ml-gcp-vertex-rag-chatbot)

Graph learning

• Fey, M., and Lenssen, J.E. "Fast Graph Representation Learning with PyTorch Geometric." ICLR Workshop, 2019. arxiv.org/abs/1903.02428 (ml-movie-recommender)
• Perozzi, B., Al-Rfou, R., and Skiena, S. "DeepWalk: Online Learning of Social Representations." KDD, 2014. arxiv.org/abs/1403.6652 (ml-social-network-predictor)
• Backstrom, L., and Kleinberg, J. "Romantic Partnerships and the Dispersion of Social Ties." CSCW, 2014. arxiv.org/abs/1310.6753 (ml-social-network-predictor)

Transfer learning and convolutional networks

• Simonyan, K., and Zisserman, A. "Very Deep Convolutional Networks for Large-Scale Image Recognition." ICLR, 2015. arxiv.org/abs/1409.1556 (ml-recyclable-material-classifier-vgg16)
• Russakovsky, O., et al. "ImageNet Large Scale Visual Recognition Challenge." IJCV, 2015. arxiv.org/abs/1409.0575 (ml-satellite-image-classifier, ml-recyclable-material-classifier-vgg16)

Clinical assessment

• Wolf, S.L., et al. "Assessing Wolf Motor Function Test as Outcome Measure for Research in Patients After Stroke." Stroke, 32(7):1635–1639, 2001. doi.org/10.1161/01.STR.32.7.1635 (ml-wearable-motion-classifier)

Datasets

• NOAA Global Historical Climatology Network Daily (GHCN-D). ncei.noaa.gov (ml-boston-climate-modeler)
• Harper, F.M., and Konstan, J.A. "The MovieLens Datasets: History and Context." ACM TIIS, 5(4):1–19, 2015. doi.org/10.1145/2827872 (ml-movie-recommender)
• Leskovec, J., and Mcauley, J. "Learning to Discover Social Circles in Ego Networks." NeurIPS, 2012. snap.stanford.edu/data/ego-Facebook.html (ml-social-network-predictor)

License

This repository is licensed under the MIT License.