Siddharth Pal सिद्धार्थ पाल

• Developing predictive models on human-skeleton data represented as spatial-temporal graphs, using dynamic graph modeling to distinguish whether a person is likely to fall or recover while walking.
• Applying explainable AI to identify the body regions and graph interactions driving fall-risk predictions, supporting interpretable analysis of human movement and stability.

• Built an AI-driven ontology mapping system on a multi-agent LangGraph pipeline (ingest, mapping, codegen, validation) combining LLMs, vector RAG, and Knowledge Graphs to automate schema alignment across heterogeneous formats (RDF/OWL, JSON, YAML), driving data interoperability and cutting manual integration effort by 65%.
• Designed an evidence-based confidence-scoring engine (patent filed) with an RLHF-inspired ML scorer trained on expert feedback, reaching 0.888 AUC-ROC and delivering explainable, transparent schema matches.
• Engineered a hybrid retrieval + multi-pass mapping pipeline (FastAPI, Qdrant, Docker) with self-healing code generation, deterministic caching, and A2A/MCP agent interoperability for reproducible, production-scale runs.
• Won Siemens Aavishkar ’25 and ’26, the company’s national Industrial AI Innovathon, placing 1st in both editions.

→ The confidence-scoring work led to a pending patent (2026)

• Investigated bias in brain-informed Large Language Models and explored techniques for its mitigation.

• Led DRDO-funded research on Hierarchical Federated Split Learning for UAV-assisted edge–fog–cloud networks, optimizing model partitioning to minimize latency.
• Validated system resilience against non-IID data distributions in dynamic environments, analyzing spatio-temporal filtering impacts on convergence.

→ This work was published in IEEE TNSM (2026)

• Automated P&ID entity extraction via OpenCV + PyTesseract; constructed Neo4j Knowledge Graph to map relationships and dependencies across industrial components.

• Fine-tuned Qwen2.5-3B with QLoRA on 2,645 execution-verified tool-calling conversations built from Walmart M5 sales data, lifting task success from 38% to 90% on 500 held-out scenarios; the agent runs locally on a MacBook via llama.cpp + Metal.
• Built a FastAPI safety gateway that validates every tool call and grounds every number in tool results, escalating only the hardest 6% to a large model: 95.4% task success, 94% answered locally, 0.4% harmful actions and 94% lower API cost, monitored with Prometheus + Grafana.

• Benchmarked PyTorch DistributedDataParallel strong scaling on 4× V100 GPUs (Slurm + torchrun): ResNet18/CIFAR-100 reached a 1.63× speedup, while a U-Net SciML surrogate got slower (0.95×); diagnosed the gap with PyTorch Profiler and NCCL kernel traces as a communication-to-compute imbalance.
• Cut gradient all-reduces per update from 12 to 3 with DDP.no_sync() gradient accumulation (+9.7% throughput, −36% NCCL time), and showed SyncBatchNorm costs 46% throughput with no accuracy gain; all results are the mean of 3 runs.

• Developed a Multi-Agent Orchestrator managing specialized AI agents for research, planning, and reasoning.
• Orchestrated 5+ specialized agents with recursive feedback loops exploring AGI pathways; integrated LLMs with LangChain for autonomous decision-making.

• Engineered an AI-driven biomedical research assistant (LangGraph + LangChain) to orchestrate MedLlama with real-time retrieval (Tavily) for complex query decomposition.
• Delivered an evidence-grounded analysis pipeline supporting parallelism and structured reasoning to generate high-throughput, research-backed summaries.

• Developed an end-to-end Text-to-STL pipeline using RAG, FAISS, and DeepSeek-V3 to synthesize high-fidelity, 3D-printable models from natural language descriptions.
• Implemented automated geometric validation (manifold checks) to ensure manufacturing reliability.

• Built a diffusion-model + U-Net pipeline for automated multi-class knee MRI segmentation (NIfTI volumes), with diffusion guidance for smooth, consistent masks across both 2D slice and 3D volume inference.
• Engineered a reproducible training workflow with class-weighted loss, TorchIO 3D augmentation, checkpointing and Dice/F1 and IoU evaluation, scaling to multi-GPU via Hugging Face Accelerate and packaged with Docker/conda.