About Sieve

Prior art, named

INGENZI Tech (MICCAI 2025 workshop) built a text chatbot over one ultrasound manual with a closed model and reported 100 percent on error codes and 80 percent on procedures with no baseline. Jua (Devpost 2023) did document Q&A with GPT-3.5. Sieve adds photo-first identification, a cited indicator registry, deterministic safety gates, an independent whole-manual audit, recall checks, a filed work order and fleet view, offline mode, and an external evaluation with baselines, including on INGENZI's own published dataset.

Where the models run

• Nebius Token Factory: Nemotron 3 Nano and 3.5 Lightning (fast tier), Nemotron 3 Super (synthesis), Nemotron 3 Ultra (audit with the whole manual in a one-million-token context).
• Nebius Token Factory: NVIDIA Nemotron 3.5 Super VL and a second image model for nameplate and panel reading, and Qwen3-Embedding-8B for hybrid retrieval over 2,994 concentrator pages.
• Laptop: manual ingestion (33 seconds for the first 9,458 pages; 9,462 indexed now) and the offline mode on a 4-bit NVIDIA Nemotron Nano 9B v2 via llama.cpp. Nebius Serverless: a Job verified both entries from a clean machine (PASS), and an Endpoint served Nemotron Nano 12B VL for a photo benchmark. Both were stopped after use.
• Tavily: manual discovery, extraction and parts lookup. openFDA: recalls and malfunction reports. Hermes Agent: chat and cron surface. NVIDIA OpenShell: sandboxed run with brokered credentials, demonstrated.

Validated against real-world evidence

Sieve is built on what biomedical technicians and hospitals have already measured: the Malawi and Gabon health-worker surveys that quantified the knowledge gap, the NEST360 survey of 63 hospitals on equipment downtime, and 647 FDA malfunction reports written by people who opened the machines. It is tested on data nobody on the team wrote: the closest prior work's own 300 repair questions, 74 pre-registered FDA reports and 57 photos of real concentrators. Every result is on the Evals page.

Feedback on the tools

Kept as a running log in the repository (FEEDBACK.md): onboarding time, what worked, what needs work, whether we would build with them again.