The Scientific Foundation Model (SFM)

Engineering intelligence
from first principles.

We are building a new class of foundation model that understands the language of engineering—physics, mathematics, and causal reasoning.

What is an SFM?

A Scientific Foundation Model (SFM) is a large-scale model trained to reason with scientific structure—equations, constraints, and empirical data—enabling physically consistent prediction, validation, and design across engineering domains.

Science-Native Learning

We integrate conservation laws, symmetries, and known constraints into training objectives to move beyond text fluency.

From Simulation to Synthesis

SFMs compress iteration cycles: rapid exploration → verifiable designs → deployable solutions, with model-assisted optimization.

Composable by Domain

Adaptable to aerospace, energy, and engineering via modular priors and calibration against domain benchmarks.

Beyond Language: The SFM Core

LLMs master language. SFMs aim to master the laws of science and engineering.

LLM: Pattern Matching

  • Trained on text & code corpora
  • Statistical correlations
  • Describes engineering concepts
  • No inherent physical guarantees

SFM: Causal Reasoning

  • Trained on simulations & equations
  • Encodes constraints (conservation, symmetry)
  • Generates verifiable designs
  • Targets physics-consistent outputs

How We Build

Rigor over Hype

Testable claims, reproducible pipelines, and honest reporting.

Compound Ownership

Small teams own literature → design → training → eval → integration.

Engineering Excellence

Brings scientists + engineers together.

Velocity with Guardrails

Move fast, measure twice—peer review and safety protocols are default.

Transparent Collaboration

  • Open code and experiment artifacts; encourage reproducibility in all projects.
  • Frequent cross-team syncs, sharing of insights, and public benchmarks.
  • Document decisions (why & trade-offs), not just the outcomes.

Responsible Innovation

  • Prioritize data privacy, bias mitigation, and verifiability in model outputs.
  • Build for energy efficiency, resource-aware design, and environmental impact.
  • Embed safety, fairness, and verification tests early in model and system design.

Open Roles

If a role is an ~80% fit and you learn quickly, we want to hear from you. Demonstrated work beats generic CVs.

Foundation Model Architecture Team

Architect large-scale transformer-based models (50B+ parameters) for India's sovereign AI initiative. Design novel attention mechanisms, tokenization strategies, and position encodings to maximize efficiency.

TransformersMoE/GQAFSDP/DeepSpeed

Data Engineering Team

Build and optimize high-throughput data pipelines for large-scale AI and SciML training. Develop systems for ingesting, cleaning, and curating scientific datasets across multiple domains.

ETLPetabyte-scaleHPC Clusters

HPC for AI Team

Architect massively parallel systems for training and inference at national supercomputing scale. Optimize performance across multi-node GPU clusters using MPI, NCCL, and custom kernels.

GPU ClustersMPI/NCCLDistributed

Scientific Computing and SciML Team

Develop physics-informed neural networks (PINNs) and Fourier Neural Operators (FNOs) for complex PDEs. Combine classical numerical methods with modern AI for multi-physics simulation.

PINNs/FNOsFEM/FVMMulti-physics

Collaborations & Advisors

We co-design benchmarks and share ablations with select research and industry partners.

ACADEMIA
INDUSTRY LABS
SYSTEM INTEGRATORS

Follow Our Progress

Request our technical brief or get notified when we release evaluations, publications, or new roles.