A single model pretrained on procedural video, endoscopic images, clinical text and medical knowledge — so detection, characterization, reasoning and reporting share one understanding of the case instead of four disconnected algorithms.
Multimodal foundation model for endoscopy
Expand endoscopy capacity. Miss fewer lesions. Cut reporting time by 80%.
One multimodal foundation model follows the whole colonoscopy: it segments and characterizes lesions frame by frame, explains every call with step-by-step Chain-of-Thought reasoning, answers your questions about what it sees, and writes the guideline-compliant report before you leave the room.
Built on medical AI research from LMU Munich.
Foundation model · live inference
Video · text · clinical data
Chain-of-Thought output
Type II pit pattern, regular vessels → low invasion risk → benign adenoma.
Clinical copilot · prototype
Built for clinical reality
Multimodal foundation model, not a single-task detector
Chain-of-Thought reasoning behind every finding
Hardware-agnostic — runs on any scope manufacturer
What sets the model apart
No black-box confidence scores. The model states what it observed, what that implies, and how it reached its classification — the same diagnostic logic a senior gastroenterologist would articulate.
Ask the copilot about any finding, in real time or afterwards: which features indicate a non-neoplastic lesion, or which resection technique is recommended — and get an evidence-backed answer.
Procedure video and metadata become a finalised, guideline-compliant report in under two minutes instead of five to ten — saving 4–8 minutes per patient and adding 15–20% daily capacity.
Demand for endoscopy is growing faster than the specialists who deliver it. We help each of them do more, and miss less.
Tandem colonoscopy studies show a 26% adenoma miss rate — above 34% for flat lesions. Detection rates range from 15% to 50% depending on who holds the scope, and every 1% drop in detection means roughly 3% more interval cancers.
The gap widens outside major academic centres, where patients face up to 70% higher interval cancer risk after adenoma removal. Gastroenterologists across Europe told us the missing piece is not another isolated prediction model, but a foundation model that follows the entire procedure, explains itself, keeps quality consistent between sites, and takes the paperwork off their desk.
Automated report generation
More procedures.
Same specialist capacity.
After a 30-minute colonoscopy, clinicians spend another 5–10 minutes dictating and drafting the report. Because the foundation model has already understood the procedure, it converts the video and metadata into a review-ready, guideline-compliant report in under two minutes — an 80% reduction in post-procedure charting time.
The economic case
Capacity is the outcome that matters.
Saving 4–8 minutes per patient sounds small. Across a high-volume endoscopy unit it compounds into 15–20% more procedures per day with the same team — without changing the examination itself.
Current physician workflow
Procedure
~30 min
Dictation and report drafting
5–10 min
Total cycle
35–40 min
With our copilot
Procedure
~30 min
Report review and approval
1–2 min
Target physician cycle
~32 min
15–20%
more procedures per day per facility, based on current workflow assumptions
and interviews with practising gastroenterologists.
Workflow assumptions reflect current product planning and conversations with practising gastroenterologists. Realised capacity gains will be measured prospectively during clinical pilots.
One copilot, end to end
See. Understand. Document.
Legacy CAD systems draw a box, show a confidence score, and stop there. Our multimodal foundation model is a single intelligence layer that follows the case from live video to the final clinical record — perceiving, reasoning, conversing and documenting.
01
See
Boundary-level segmentation of polyp contours on the live video feed — including flat and sessile serrated lesions and small polyps behind mucosal folds. Each finding is classified against established criteria such as Paris, NICE and BASIC.
02
Understand
Chain-of-Thought reasoning replaces the unexplained alert: what the model observed, what it implies, and how the lesion was classified. Through Visual Question Answering you can interrogate any finding during or after the procedure — which features suggest it is non-neoplastic, or which resection technique is recommended — and receive an evidence-backed answer.
03
Document
The procedure video becomes a finalised structured report in under two minutes: bowel preparation score, polyp count, sizes, anatomical locations and surveillance recommendation — an 80% cut in charting time, ready for clinician sign-off and integration with the hospital record via HL7/FHIR.
Beyond computer-aided detection
From finding a polyp to understanding the procedure.
Lesion-level intelligence
Bounding boxes and confidence scores on individual frames, tied to one manufacturer's hardware. The clinician interprets each alert, and the documentation stays manual.
Procedure-level intelligence
One multimodal foundation model that follows the procedure over time, explains its reasoning step by step, answers clinical questions through VQA, tracks procedural quality and writes the report — on any scope manufacturer's equipment.
Long-term vision
Colonoscopy is the starting point.
A multimodal foundation model that already combines procedural video with images, text and clinical knowledge is not limited to one organ or one task. Colonoscopy is where we prove it; clinical video across specialties is where it goes.
Colonoscopy
Detection, characterization, procedure understanding and automated reporting.
Gastrointestinal endoscopy
Extend the same procedure intelligence to upper GI and other endoscopic workflows.
Interventional care
Bring multimodal procedural understanding into image-guided specialties and surgery.
Clinical video infrastructure
A model-agnostic layer for understanding, reasoning, reporting and integration with hospital systems.
The team
Research depth. Clinical ambition.
We combine frontier medical AI research, hands-on experience building certified medical AI products, and the financial background to bring them to market.
Mina Rezaei, Ph.D.
Mina leads research and AI. She heads an AI research group and lectures at LMU Munich, where she has led interdisciplinary projects in trustworthy AI, foundation models, medical imaging and uncertainty quantification. She has published more than 80 papers at venues including NeurIPS, ICLR, MICCAI and Nature, and has supervised six PhD students and over 60 master's and bachelor's theses.
Shunjie-Fabian Zheng, M.Sc.
Fabian leads business and strategy. He combines hands-on medical AI experience from Google Brain and Lunit's cancer screening group with corporate finance experience from investment banking and mid-market private equity — the mix needed to take a clinical AI product from research to a regulated market.
Contact
Let’s build the future of clinical video.
We are speaking with hospitals and endoscopy centres about pilots, and with research partners and investors who want to advance procedure-level AI in healthcare.
Co-Founder · Research & AI
Mina Rezaei, Ph.D.
mi.rezaei11@gmail.com
Co-Founder · Business & StrategyShunjie-Fabian Zheng, M.Sc.
fabian_zheng@yahoo.de
Questions? Write to either of us.
Building in Munich, Germany
Colonoscopy is the start. Clinical video is the goal.
Our foundation model is moving from research prototype to clinical pilots. We are welcoming hospital partners for data access, validation and prospective deployment.
Get in touch