CausalForge

Inspiration

The core insight came from a frustration with how LLMs handle science: they summarize, they hallucinate, and they forget the graph structure of causality. A paper doesn't just say things — it asserts directed causal relationships with specific strengths, signs, and uncertainty bounds.

The question I wanted to answer: what if instead of summarizing papers, you compiled them into an executable causal program?

Challenges we ran into

Structured extraction at scale: Getting NIM to reliably output valid JSON claim objects required careful prompt engineering with strict schema enforcement and retry logic.

Graph consistency: Merging claims from many papers into a single coherent graph without duplication or contradictory overwrites required a canonical variable normalization step.

Simulation fidelity vs. speed: Full probabilistic Bayesian propagation is computationally expensive. The current engine uses a deterministic approximation (signed BFS with confidence decay) that is fast and inspectable.

What we learned

Causal graphs are a fundamentally better substrate for scientific knowledge than embedding vectors — they are inspectable, simulable, and falsifiable.

NVIDIA NIM's structured output capabilities are powerful enough to serve as a scientific parsing layer with the right prompting strategy.

MongoDB Atlas's document model is a natural fit for heterogeneous claim objects where schema varies by paper domain.

Built With

• fastapi
• mongodb
• next.js
• nvidia
• python
• tailwind