Pankhuri Kulshrestha
Machine Learning Research Engineer at Mindtrace, Manchester — building distributed monitoring agents for industrial inspection workflows, with research spanning multi-agent coordination, federated learning, and 3D deep learning. My core question: how do you build systems that learn continuously, privately, and reliably in the real world.
Recent Achievements
| Apr 2025 | Published at SAIA 2025 (IEEE) — Evaluating Dialogue Adaptability: Self-Feeding Mechanisms in Federated and Centralized Chatbot Architectures |
| Apr 2025 | Published at IEMTRONICS 2025 (Springer) — Privacy-Preserving Self-Learning Chatbot with Federated Intelligence |
| 2025 | Designing the agents package for Mindtrace — autonomous monitoring agents for distributed weld inspection workflows |
| 2023–Present | 3D deep learning and MLOps at Mindtrace — LiDAR point cloud segmentation, DINO architectures/YOLO defect detection (92% accuracy), and production infrastructure (Docker, MLFlow, FastAPI) for industrial asset inspection and manufacturing defect detection |
Selected Work
Mindtrace — Open-Source ML Infrastructure Framework Contributor · Core Package & Agents Designed the monorepo architecture and core package (config injection, structured logging, observable state); building the agents package for distributed weld inspection — coordinating tasks, maintaining memory across sessions, and surfacing anomalies to operators.
ML Deployment Infrastructure Production MLOps · IQVIA Reusable Docker + Kubernetes + Helm templates for packaging and serving ML models as microservices — persistent volume mounting, FastAPI inference scaffolding, and CI/CD hooks (GitLab CI, Jenkins). Distilled from production deployment work across AWS, GCP, and Azure.
Privacy-Preserving Self-Learning Chatbot with Federated Intelligence IEMTRONICS 2025 · Springer Federated learning framework enabling chatbots to learn from user interactions without centralising private data — keeping sensitive conversations on-device throughout training.
Evaluating Dialogue Adaptability: Self-Feeding Mechanisms in Federated and Centralized Chatbot Architectures SAIA 2025 · IEEE Comparative study of federated vs. centralised training dynamics — analysing convergence, adaptability, and where the privacy-performance trade-off bites.
Autonomous Robot Navigation with ROS2 MSc Coursework · University of Essex ROS2-based autonomous navigation system using LIDAR sensing, a 5-stage state machine, and dual control strategies — PID for wall-following and fuzzy logic for gap navigation — benchmarked through a maze environment.
Research Interests
- Distributed Multi-Agent Lifelong Learning
- Agent Memory & Collective Knowledge
- Agentic RAG & Cache-Augmented Generation
- Long-Horizon Planning & Self-Reflection
- Computer Vision & 3D Deep Learning (LiDAR, PointNet, DINO architectures, SAM 2)
- Federated Learning & Privacy-Preserving ML
- Agent Safety & Alignment