AI-Based Web App for Scientific ML Inferences
Client: Accelerate Scientific Simulations using Integrated AI
Overview
A UAE-based company focused on building high-fidelity specialized ML models for scientific research and discovery approached Travancore Analytics with a unique request. They needed a web-based, fully managed platform that accelerates physics-driven scientific computing by integrating advanced ML models with simulation workflows.
We designed,
- a browser-accessible platform that enables users to leverage a curated library of ready-to-use, pre-configured, physics-informed models.
- an AI chatbot for simulation control using NLP, and
- real-time visualization capabilities, all with a zero local setup.
The platform was targeted towards scientists and engineering professionals who needed reliable, reproducible results for complex engineering and scientific problems, leading to lower time-to-inference and operational overhead.
The Case
Traditional LLM solutions are programmed to be conversational and are not suited for handling complex physics-based queries. Researchers are forced to use multiple tools and still have to face an ML/DevOps domain expertise barrier. Most scientific ML toolkits require local GPUs or cluster provisioning, which also increases both cost and complexity substantially. The setup cycle of such scientific ML models takes a considerable amount of time with the GPU (CUDA) configuration and dependency resolutions, which ultimately delays experimentation. It also required a high computational cost to run iterative research in traditional high-fidelity simulation models.
Challenges
Bitwise Reproducibility: Ensuring 100% reproducibility across model runs with the same inputs and configurations while hiding dependency chains from the end user
Numerically-Stable Inference: Delivering high-performance, numerically-stable inference for a diverse set of ML models in a browser-accessible service.
Accuracy: Maintaining faster time-to-inference without compromising on accuracy such that ML-augmented results remain sound enough to be used in scientific research.
Technology Barrier: Guaranteeing that the UX supports interactive exploration and batch workflows for non-ML experts
Simplifying simulation control: Integrating real-time visualization and simulation control without exposing users to the underlying computational complexity.
Solution
Key features
Technologies used
Python, React, FastAPI, AWS, LangChain, LangGraph, Groq
Curated Model Library: Around 10 physics-informed models ready for immediate use.
Zero-Setup Inference: Local setup not required
AI-Powered Chat Interface: Natural language control for parameterization and simulation steering.
Real-time Visualization: Dynamic plotting and playback for simulation outputs.
Impact
Get in Touch
Ready to Transform
Your Business?
Send us a Message
We’ll respond within 24 hours