Projects

Challenges

• — Tackling extreme class imbalance (0.17% fraud prevalence across 284,807 transactions) where standard cross-entropy and accuracy metrics fail
• — Minimizing Expected Calibration Error (ECE < 0.0002) via Platt Sigmoid Scaling to ensure predicted probabilities match empirical risk
• — Balancing severe asymmetric business costs: preventing chargeback loss while eliminating 96.8% of false positive review overhead ($50/case)
• — Meeting strict sub-millisecond inference SLAs (<1.0ms) with high-throughput batch stream processing (139k+ tx/sec)

Learnings

• — Engineered a Platt-calibrated soft-voting meta-ensemble achieving 0.8384 PR-AUC and 96.77% Precision on 42,722 unseen test transactions (only 2 false positives)
• — Designed a custom PyTorch Deep Tabular ResNet with skip-connections and Binary Focal Loss (γ=2.0) to suppress easy-negative gradient dominance
• — Built an asynchronous FastAPI microservice with automated schema validation and dynamic localhost collision resolution
• — Developed an interactive minimalist obsidian glassmorphic dashboard with autonomous client Edge ML simulation for zero-downtime static deployments

Challenges

• — Designing a stateful multi-agent optimization loop that can autonomously plan, execute, evaluate, and iterate on ML experiments
• — Ensuring model improvements are statistically meaningful using paired bootstrap confidence intervals, McNemar's test, and effect-size analysis
• — Building a single MCP tool implementation usable by both in-process LangGraph agents and external MCP clients
• — Safely handling agent-driven code execution with command allowlisting, workspace isolation, environment sanitization, timeouts, and output limits

Learnings

• — Built a five-agent LangGraph workflow covering planning, experimentation, statistical evaluation, optimization, and deployment
• — Implemented MCP-based tooling for dataset inspection, model training, statistical testing, model registration, deployment, rollback, and monitoring
• — Integrated MLflow experiment tracking, versioned model registry, FastAPI serving, Prometheus metrics, and PSI-based drift monitoring
• — Created a deterministic offline mock-LLM mode that makes the complete agent workflow reproducible without an API key

Challenges

• — Reproducing MSE-optimal quantization using random rotations and Lloyd-Max scalar quantization
• — Integrating MSE quantization with residual QJL correction for inner-product-preserving compression

Learnings

• — Implemented both core TurboQuant algorithms in Python from the paper
• — Built TurboQuantProd combining MSE quantization with residual QJL correction

Challenges

• — Building an end-to-end pipeline spanning preprocessing, feature engineering, training, and evaluation
• — Ensuring prediction quality across MAE, RMSE, and MSE metrics

Learnings

• — Achieved an R² score of approximately 0.74 on revenue forecasting
• — Applied SHAP feature attribution to identify key business drivers for revenue