§ boringai · image pipeline

Three-stage fallback chain.

Every image call escalates through three models on Modal GPU. The first one that produces trustworthy output wins. MedGemma is the medical-specialized primary; LLaVA is the general-vision fallback; BLIP + clinical LLM is the safety net that never fabricates.

IMAGE + PROMPT Reviewer-gated medical output returned to the caller · clinical sign-off required STAGE 01 MedGemma 4B IT google/medgemma-4b-it 4.3B params · bf16 ✓ DIRECT ANALYSIS × STAGE 02 LLaVA-1.5-7B llava-hf/llava-1.5-7b-hf 7B params · fp16 · non-gated ✓ DIRECT × STAGE 03 BLIP Salesforce/blip-image-captioning-base 250M params · safety net → INTERPRET Caption + clinical LLM glm-4.7-flash honest when caption is junk ✓ INTERPRETED LEGEND success — use output directly fail — try next stage interpret via clinical LLM PIPELINE first success wins · every stage wraps in try/except · guards reject <16px and blank images
boracode.ai · boringai medical inference deployed: modal / boringai-medical