Machine Learning Projects Portfolio

A collection of production-grade machine learning implementations built from first principles. Each project demonstrates deep understanding of core algorithms through clean, modular code, reproducible validation, and interview-ready documentation.

Projects Overview

0. Real-Time Model Serving and Monitoring Pipeline

Directory: real-time-model-serving-monitoring/

A production-style MLOps project demonstrating real-time inference, model artifact versioning, request validation, prediction logging, latency telemetry, and feature-drift monitoring.

Key Features:

• FastAPI service with /health, /predict, /batch_predict, /monitor/drift, and /monitor/latency
• Reproducible scikit-learn training pipeline with generated model artifact, metadata, baseline feature statistics, and model card
• p50/p95 latency tracking, request logging, and PSI-based feature drift monitoring
• Dockerfile, example payloads, and automated tests covering API and monitoring paths
• Holdout metrics from reproducible run: Accuracy 94.7%, ROC-AUC 0.993, F1 0.958

Technologies: FastAPI, Docker, scikit-learn, Pydantic, pytest, model monitoring

Use Cases: MLOps, production ML, real-time inference, model monitoring, AI/ML engineer interviews

---

1. Diffusion Models from Scratch

Directory: diffusion-models/

A complete implementation of Denoising Diffusion Probabilistic Models (DDPM) for image generation.

Key Features:

• Forward and reverse diffusion processes with multiple noise schedules
• Custom UNet architecture with attention mechanisms
• DDPM and DDIM sampling algorithms
• Classifier-free guidance for conditional generation
• Trained on CIFAR-10 dataset
• FID reporting and ablation workflow

Technologies: PyTorch, CIFAR-10, UNet, attention mechanisms

Use Cases: Image generation, denoising, creative AI applications

---

2. Mini-GPT: Decoder-Only Transformer

Directory: mini-gpt/

A GPT-style language model implementing the complete transformer architecture from scratch.

Key Features:

• Multi-head self-attention with causal masking
• Pre-LayerNorm transformer blocks
• KV cache for efficient autoregressive generation
• Mixed precision training
• Multiple sampling strategies: greedy, top-k, top-p, temperature
• Configurations from small to medium-scale experiments

Technologies: PyTorch, Transformers, Attention, NLP

Use Cases: Text generation, language modeling, conversational AI

---

3. PPO from Scratch

Directory: ppo-reinforcement-learning/

A complete implementation of Proximal Policy Optimization for reinforcement learning.

Key Features:

• Clipped surrogate objective for stable learning
• Generalized Advantage Estimation
• Separate actor-critic networks
• Custom inventory management environment
• Trained on CartPole and custom supply-chain optimization
• Current inventory result uses random-policy baseline; classical EOQ / newsvendor / (s,S) baselines should be added before resume use

Technologies: PyTorch, Gymnasium, Reinforcement Learning, Policy Gradients

Use Cases: Operations research, inventory control, RL methodology

---

4. Mistral 7B LoRA / QLoRA Fine-Tuning

Directory: mistral-7b-qlora-lora-finetuning/

Instruction fine-tuning project currently under active training. It should be used on resumes only after training completes, evaluation is reproducible, and LoRA hyperparameter ablations are documented.

Target Evidence Before Resume Use:

• Base vs fine-tuned evaluation
• LoRA rank / alpha / dropout ablations
• QLoRA memory and throughput report
• Held-out instruction-following evaluation
• Clear limitations around dataset quality and hallucination risk

---

5. Credit Risk and Customer Lifecycle Analytics

Directory: credit-risk-customer-lifecycle-analytics/

End-to-end fintech data science project for credit risk, banking analytics, customer lifecycle, scorecards, and portfolio monitoring roles.

Key Features:

• Synthetic but realistic 25,000-customer banking portfolio with bureau, repayment, utilization, acquisition, churn, and margin signals
• Temporal validation by customer vintage month to avoid random-split leakage
• Model comparison across logistic scorecard, gradient boosting, and random forest
• ROC-AUC 0.763, KS 0.405, and 34.5% recall in the riskiest decile from the synthetic lifecycle pipeline
• Real OpenML German Credit benchmark: ROC-AUC 0.801 and KS 0.494 on 1,000 public records
• DuckDB SQL analytics layer for approval policy, realized default, churn, segment profitability, and acquisition-channel quality
• Permutation explainability, PSI drift monitoring, model card, executive brief, plots, and tests

Technologies: Python, pandas, scikit-learn, DuckDB, matplotlib, pytest

Use Cases: Credit risk, fintech analytics, data scientist, product analytics, banking DS, model governance

---

Project Structure

ML-PROJECTS/
- real-time-model-serving-monitoring/ # FastAPI model serving + monitoring
- diffusion-models/                   # DDPM image generation
- mini-gpt/                           # GPT-style language model
- ppo-reinforcement-learning/         # PPO RL implementation
- mistral-7b-qlora-lora-finetuning/   # Active LoRA / QLoRA fine-tuning work
- credit-risk-customer-lifecycle-analytics/ # Credit risk + customer lifecycle analytics
- README.md                           # This file

Common Characteristics

All projects aim to demonstrate:

• From First Principles: Core algorithms implemented explicitly where educational value matters
• Production Quality: Modular design, documentation, error handling, and tests where appropriate
• Research Grade: Clear baselines, reproducible metrics, ablations, and honest limitations
• Interview Readiness: Each project should have defensible bullets, caveats, and walkthroughs

Technologies Used

• Deep Learning: PyTorch, Transformers, diffusion models, reinforcement learning
• MLOps: FastAPI, Docker, Pydantic, pytest, model monitoring
• Python Stack: NumPy, pandas, scikit-learn, matplotlib, TensorBoard where relevant
• Environments: CPU-friendly tests by default; GPU required only for heavier training projects

Getting Started

Each project has its own README with detailed instructions.

# Real-time model serving
cd real-time-model-serving-monitoring/
pip install -r requirements.txt
python scripts/train_model.py
python -m pytest -q
uvicorn src.ml_monitoring.app:app --host 0.0.0.0 --port 8000

# Diffusion Models
cd diffusion-models/
pip install -r requirements.txt
python training/train_model.py --config configs/base.yaml --experiment baseline

# Mini-GPT
cd mini-gpt/
pip install -r requirements.txt
python training/train_model.py --config configs/small.yaml

# PPO
cd ppo-reinforcement-learning/
pip install -r requirements.txt
python training/train_agent.py --env CartPole-v1 --total-timesteps 50000

Resume Use

For AI/ML Engineer resumes, prioritize:

1. Real-Time Model Serving and Monitoring Pipeline
2. Mini-GPT
3. DDPM / diffusion project
4. Mistral LoRA / QLoRA only after training and ablation results are complete

FlowFinance production AI work should be listed as venture / founder experience, not as a project.

For Data Scientist / fintech analytics resumes, prioritize:

1. Credit Risk and Customer Lifecycle Analytics
2. SQL Operations Analytics
3. Demand Forecasting + Safety Stock
4. A/B Testing and Funnel Analytics

License

Each project is individually licensed. See the respective license files in each project directory.