Prism.ai

Prism.ai was built for OpenAI Build Week 2026. It answers a narrow question: how much of the mechanical work of an empirical research cycle can be automated without letting the model invent results?

The pipeline

The system runs six stages, each producing an artifact the next stage consumes:

1. Dataset analysis: deterministic profiling of the uploaded dataset: shape, class balance, missingness, leakage checks, and for image data a set of computed vision statistics. No model is involved here, so the numbers downstream stages reason about are measured rather than guessed.
2. Literature retrieval: queries the OpenAlex corpus for work relevant to the dataset and task, then filters and ranks candidates for methodological relevance instead of citation count alone.
3. Experiment planning: proposes a concrete experimental design: baselines, ablations, metrics, and the splits each will be evaluated on.
4. Code generation: emits runnable PyTorch training and evaluation code for the plan, with the data-loading layer bound to the profile from stage one.
5. Execution and analysis: runs the experiments and collects metrics.
6. Research report: assembles the measured results, the retrieved literature, and the experiment plan into a structured report with the provenance of every number preserved.

Design decisions

The central constraint was separating measurement from generation. Anything that could be computed deterministically (dataset statistics, metric calculation, and split construction) is computed, not prompted. The language model is used for the parts that are genuinely generative: framing the question, selecting relevant prior work, proposing designs, and writing prose. Every quantitative claim in the final report traces back to a stage that computed it.

The orchestration layer is FastAPI, the training and evaluation code targets PyTorch, and the interface is a Next.js app that streams each stage's output as it completes so a long-running pipeline stays legible while it works.