Aaditya Bhatnagar

◦ Built an end-to-end Spanish power-market forecasting platform using PostgreSQL and Parquet for electricity price, demand, generation, and weather data.
◦ Engineered leakage-safe time-series features and compared baseline, Ridge, random forest, LightGBM, and
gradient-boosting models with temporal splits, MLflow tracking, and walk-forward validation.
◦ Converted forecasts into simple long, short, or flat trading signals and tested them with transaction costs, risk metrics, FastAPI endpoints, React dashboards, and an LLM assistant.

◦ Built a 2-4-1 MLP visualizer with core logic in Rust, implementing forward/backpropagation, MSE loss, SGD, seed value for deterministic initialization, and support for XOR, half-moons, and spiral datasets.
◦ Exposed training and inference through an Axum/Tokio control plane and React interface for live weights,
activations, loss curves, and predictions. Added an agentic FastAPI tutor layer using Groq tool calling, bounded multi-turn memory, SSE streaming, MCP integration, Langfuse/OpenTelemetry tracing, and containerized Rust, Python, and React services.

◦ Built an end-to-end B2B support copilot with multi-format document ingestion, Qdrant-backed hybrid dense/BM25 retrieval, cited Gemini-generated responses, PostgreSQL answer traces, and safe escalation for unsupported queries.
◦ Developed labelled evaluation pipelines for retrieval and grounded response behavior, measuring citation grounding, expected answer-term coverage, and unsupported-query handling.
◦ Achieved 100% top-citation accuracy, 93.8% expected answer-term coverage, and 100% unsupported-query escalation accuracy on a product-specific benchmark.