SubmissionAI is a long-horizon review harness for evolving drug-development document packages. It stores immutable package versions, evidence passages, reviewer-approved memory, review history, and workflow checkpoints in the MongoDB Atlas Sandbox. Nebius models power review and embeddings, while Atlas Vector Search retrieves passages scoped to the active study, package, and embedding profile. For each revision, the harness fingerprints source dependencies and model/retrieval settings. It reruns affected checks and reuses unchanged, verified results. Findings can resolve when documents align or reopen if a later version restores a conflict. Model citations and conclusions are checked against current source text and explicit review rules; saved checkpoints support resuming interrupted work. The synthetic v1→v2→v3 demo shows a timing mismatch and missing appendix, a corrected version that reuses the unaffected coverage check, and a later version that reopens the timing finding. In one paired synthetic rehearsal, incremental review used one model call versus three for a full rerun, with the same verified findings. This prototype is not medical or regulatory advice.
Team
Placement
Participant
