Machine Learning 10 projects
Classic algorithms, training & tuning
Customer Churn
Flags telco subscribers likely to cancel; compares LogReg / RF / XGBoost + a stacking ensemble, reports ROC-AUC and the top churn drivers.
View code & docs →House Prices
Price regression comparing Ridge / RF / Gradient Boosting + stacking; reports R²/MAE/RMSE and the features that drive price.
View code & docs →Fraud Detection
Imbalanced credit-card fraud with SMOTE oversampling + multi-algorithm stacking, tuned for recall.
View code & docs →Credit Risk
Loan-default classifier (LogReg / RF / GBM + stacking) with feature-importance explanations.
View code & docs →Breast-Cancer Diagnosis
Explainable diagnosis on UCI Wisconsin; multi-algorithm + stacking with clinician-checkable importances.
View code & docs →Customer Segmentation
RFM features from raw transactions; KMeans vs Agglomerative vs DBSCAN scored by silhouette, labelled segments.
View code & docs →Demand Forecasting
Forecasting as supervised regression on lag + calendar features vs a seasonal-naive baseline; MAE/RMSE/MAPE.
View code & docs →Stock Movement
Technical features → chronological split (no look-ahead leakage) → honest backtest of long/flat vs buy-and-hold.
View code & docs →Movie Recommender
SVD matrix factorization vs mean baselines on MovieLens, held-out RMSE + top-k recommendations.
View code & docs →AutoML + Optuna
Capstone: an Optuna study that tunes a gradient-boosting model, with a default-vs-tuned comparison and an optimization-history chart.
View code & docs →Deep Learning 10 projects
PyTorch: training, transfer learning, fine-tuning
CNN from Scratch
A SimpleCNN + training loop on Fashion-MNIST that classifies example images.
View code & docs →Transfer Learning
Swaps ResNet-18's head, freezes the backbone, fine-tunes on any ImageFolder dataset.
View code & docs →Pneumonia X-ray
Fine-tuned ResNet-18 (normal vs pneumonia) + a from-scratch Grad-CAM overlay for explainability.
View code & docs →Object Detection
A YOLO11 detector: upload an image and get back an annotated image plus a table of detections.
View code & docs →U-Net Segmentation
A compact U-Net with skip connections, pixel-wise training loop, input-vs-mask visualisation.
View code & docs →Traffic Signs
A colour-image CNN over GTSRB 32×32 crops, trained and evaluated in the chat UI.
View code & docs →LSTM Forecasting
An LSTM that predicts the next value from a window, then forecasts multiple steps recursively.
View code & docs →Speech Emotion
Audio → MFCC features → MLP classifier (librosa, with a numpy spectral fallback so it runs offline).
View code & docs →Face Verification
A CNN that maps a face to an L2-normalised embedding, then verifies pairs by cosine similarity.
View code & docs →ViT Fine-Tune
Capstone: a Vision Transformer with a swapped head, a fine-tune loop, and optional Weights & Biases sweeps.
View code & docs →NLP 10 projects
Classic NLP → transformers → LLM fine-tuning
SMS Spam
TF-IDF comparing Naive Bayes / LogReg / Linear SVM; paste a message, get spam/ham + confidence.
View code & docs →Sentiment
TF-IDF classifier comparison that scores any review; documents the DistilBERT fine-tune path for SOTA.
View code & docs →Fake News
TF-IDF NB / LogReg / SVM classifier that scores any headline.
View code & docs →Named Entity Recognition
A rule-based extractor in the chat UI; documents the fine-tuned BERT / spaCy production path.
View code & docs →Summarization
A faithful extractive summarizer (frequency-scored sentences) + the documented BART abstractive path.
View code & docs →Topic Modeling
Compares NMF (on TF-IDF) and LDA (on counts), shows top words per topic, matches new docs to a topic.
View code & docs →Machine Translation
A chat translator: an offline word-lookup demo plus the real NLLB-200 path via Transformers.
View code & docs →Question Answering
A TF-IDF sentence-relevance baseline over your context + the SQuAD BERT span-extraction path.
View code & docs →LLM Fine-Tune (LoRA/QLoRA)
16 GB GPU: Alpaca-format instruction-data prep plus a complete Unsloth QLoRA training script.
View code & docs →Semantic Search
A TF-IDF résumé ↔ job matcher + the sentence-transformer dense-embedding upgrade.
View code & docs →Generative AI 10 projects
LLM apps, RAG, image/video, local inference, guardrails
Prompt Playground
Multi-provider chat (OpenAI / Anthropic / Ollama / offline mock) that shows which model answered + a chain-of-thought transform.
View code & docs →RAG: Chat with Docs
Chunk → retrieve → answer with citations over your text or PDFs.
View code & docs →GraphRAG
Entity extraction → a networkx co-occurrence graph → neighbourhood retrieval → grounded answer.
View code & docs →Local LLM Chat
Chats against a local model via Ollama or a vLLM OpenAI-compatible server, with history.
View code & docs →1-bit LLMs (BitNet)
From-scratch ternary (absmean) quantization: it converts weights to {−1,0,1} and reports the compression and the error.
View code & docs →Image Generation
Text-to-image: a procedural offline demo plus the real diffusers path (SD 3.5 / FLUX).
View code & docs →Video Generation
Text-to-video: a procedural GIF offline demo plus the diffusers LTX-Video and hosted-API path.
View code & docs →Document Intelligence / OCR
scan/PDF → OCR → field extraction; extraction tested offline, modern OCR engines documented.
View code & docs →Structured Extraction
A Pydantic schema validated on the way in: regex offline, PydanticAI (LLM fills the schema) in production.
View code & docs →Evals + Guardrails
LLMOps capstone: quality metrics (exact-match, faithfulness, recall) and guardrails (PII, banned content), all pulled into one report.
View code & docs →AI Agents 13 projects
Frameworks & real agent types
ReAct from Scratch
A ReAct executor with a safe calculator + lookup tool and a transparent Thought/Action/Observation trace.
View code & docs →SQL Agent
Natural language → SQL over a real SQLite DB; LangGraph + a schema-aware LLM in production.
View code & docs →Web / Browser Agent
Navigates a site and extracts answers with a visible trace; browser-use / Playwright on a live browser in production.
View code & docs →Research Agent
Searches a corpus, ranks sources, writes a cited synthesis; CrewAI + a web-search tool in production.
View code & docs →Multi-Agent Content Team
A researcher → writer → editor pipeline that turns a topic into an article, showing each role's output.
View code & docs →Conversational Multi-Agent
A Solver ↔ Critic dialogue that refines an answer over rounds, with the full transcript.
View code & docs →Type-safe Tool Agent
Tools with Pydantic arg models + validators and a router that validates before executing.
View code & docs →A2A: Across Frameworks
An A2A message schema + agent cards; two cross-framework agents collaborate on one task.
View code & docs →Healthcare Agent
Guideline-RAG wrapped in safety rails: emergency escalation, refusal to diagnose/prescribe, disclaimer on every answer.
View code & docs →Finance Agent
A finance analyst over a financials table (SQLite) with a no-advice guardrail; RAG over 10-Ks in production.
View code & docs →Voice Agent
An STT → LLM → TTS pipeline; the text orchestration is tested offline, the audio edges are documented.
View code & docs →Guardrails / Safety
An input rail (jailbreak / disallowed-request blocking) + an output rail (PII masking) wrapping a model.
View code & docs →Agent Harness
Capstone: routes a task to the right skill, retries, and returns a full trace plus metrics; Claude Agent SDK or LangGraph with a tracer in production.
View code & docs →Capstone Products 10 projects
End-to-end, one per domain · full build briefs
Clinical Decision Support & Triage
Guideline-grounded triage and Q&A with hard safety rails, an audit log, and a clinician review queue, in a private cloud.
View full brief →Banking Support & Fraud Assistant
Policy Q&A plus a SQL agent over transactions, dispute intake, fraud flagging, and a no-advice guardrail.
View full brief →Adaptive Tutor & Course Builder
A Socratic tutor that hints without leaking answers, auto-builds quizzes and lesson plans, and grades with rubrics.
View full brief →Recruitment & People-Ops Copilot
Bias-aware resume-to-role matching, interview kits, an employee policy bot, and a recruiter dashboard.
View full brief →Property Discovery & Deal Analyzer
Natural-language and geo property search, vision tagging, contract extraction, and a deal-analysis agent.
View full brief →Enterprise Knowledge Search
Hybrid keyword and vector search across Drive, Confluence, and Slack, with permission-aware GraphRAG.
View full brief →No-Code AI Agent Builder
A no-code visual agent builder on Copilot Studio and Azure AI Foundry, with governance and built-in evals.
View full brief →Azure Document & Call Intelligence
Built only on Azure: Document Intelligence, Speech, OpenAI, and AI Search, with Content Safety guardrails.
View full brief →Insurance Policy & Claims Assistant
Policy Q&A with clause citations, OCR claims intake, a coverage rules engine, and an adjuster dashboard.
View full brief →Investment Research & Portfolio
RAG over filings plus a SQL agent over holdings, risk analysis, and a strict no-advice guardrail.
View full brief →