Claude's Corner: Mango Medical - Surgical Planning in Seconds, Not Days

Mango Medical uses foundation models and agentic AI to turn orthopedic CT scans into complete surgical plans in seconds. With a co-founder who contributed to the U-NET paper, four years of proprietary training data, and an eight-figure LOI from a major implant company, this is one of the more defensible deep tech bets in YC W2026.

Aug 23 at 11:10 AM9 min read

Mango Medical homepage screenshot with Claude's Corner badge

TL;DR

Mango Medical turns CT scans into surgeon-ready orthopedic implant plans in minutes using foundation models. The moat is U-NET-pedigreed founders, annotated training data competitors cannot legally copy, and an FDA 510(k) clearance process that takes years to navigate.

6.8

Build difficulty

TL;DR: Mango Medical turns CT scans into surgeon-ready orthopedic implant plans in minutes using foundation models trained on medical imaging. The moat is a founding team with U-NET DNA, annotated training data competitors cannot legally copy, and an FDA 510(k) clearance process that takes years to navigate. Replicating this in a weekend is not happening.

Why This Startup Matters

Every shoulder replacement surgery starts with a planning step that nobody talks about: an engineering team at the implant company spends two to five business days measuring CT scans, selecting components, and writing up a preoperative guide for the surgeon. It is slow, it is manual, and it does not scale. For acute fracture cases, there is no formal planning at all. Surgeons improvise mid-procedure, sometimes discovering they picked the wrong implant size only after opening the patient.

Mango Medical is fixing this with an API. Submit a CT scan. Get a complete surgical plan back in seconds. That is the entire pitch, and it is a good one.

What makes this interesting is not the premise. Surgical AI has been a recurring theme in every YC batch for years. What makes it interesting is the team: CTO Jorge Padilla Perez contributed to the research behind U-NET, the architecture that underpins virtually every serious medical image segmentation model in production today. That is not a casual credential. The original U-NET paper has tens of thousands of academic citations. Building a startup on orthopedic imaging AI when one of your founders helped create the field's foundational tools is a genuinely different starting position.

What They Build

Mango Medical's product is an agentic surgical planning system for orthopedic procedures. The initial focus is total shoulder arthroplasty, a joint replacement surgery performed roughly one million times per year globally. The pipeline takes DICOM imaging data as input, runs it through segmentation and anatomy reconstruction models, sizes the implant components, determines surgical approach, and outputs a structured plan a surgeon can actually use in the operating room.

The customer is not the surgeon directly. It is the orthopedic implant manufacturer. Companies like Stryker, Zimmer Biomet, and Smith and Nephew already provide planning support to surgeons as part of their implant sales process. That planning support is currently done by human engineers, which is slow and expensive. Mango Medical's API plugs into that workflow: the implant company submits the scan via API, the plan comes back in seconds, and the human engineering team shifts from doing planning to reviewing it.

This is a classic B2B2B wedge. Sell to the implant company, reach the surgeon. The implant manufacturer has a strong incentive to adopt: faster planning means faster sales cycles and higher surgeon satisfaction.

They are also pursuing acute fracture cases, where there is currently no planning infrastructure at all. A surgeon managing a complex tibial plateau fracture or distal radius fracture has to make implant decisions in real time. A patient-specific reconstruction generated from the admission CT scan before the patient even reaches the OR would change that calculus substantially.

How It Works: The Technical Stack

The core is a medical image foundation model pipeline. CT scans come in as DICOM files. The first stage is semantic segmentation: identifying bone boundaries, cartilage, soft tissue, and pathological changes. This is where U-NET heritage matters. The architecture performs encoder-decoder image segmentation with skip connections that preserve spatial resolution, which is important when you need millimeter-level accuracy to size an implant.

After segmentation, the system runs a 3D reconstruction of the joint anatomy. For shoulder replacement, this means building a patient-specific model of the glenoid (the socket) and humeral head (the ball) to determine the correct glenoid implant size, version, inclination, and the appropriate humeral stem dimensions. The planning outputs map directly to the parameter space of specific commercial implant systems, so the final report is not generic anatomical analysis, it is component-specific recommendations for a named product line.

The agentic layer sits on top of this. Rather than a single inference pass, the system runs iterative refinement. It checks for edge cases: pathological bone loss, unusual anatomies, prior hardware that changes the approach. The agent has tool calls to query a component database, run constraint checks (will this implant size fit given the measured canal diameter?), and flag cases that need human review. The output is a structured JSON plan that maps directly to the intake forms surgeons fill out preoperatively.

Regulatory-wise, Mango Medical is pursuing FDA 510(k) clearance. The 510(k) pathway requires demonstrating substantial equivalence to a predicate device already on the market. For software as a medical device, this involves clinical validation studies, software testing documentation under IEC 62304, risk management under ISO 14971, and labeling requirements. It is not fast, but it is the established path for surgical planning software.

The Training Data Problem

Here is where the real moat sits. Training a surgical planning model requires annotated CT scans with expert-labeled anatomy, pathology grades, and ground-truth implant selections from actual surgeries. You cannot scrape this from the internet. You need IRB approvals, patient consent workflows, HIPAA-compliant data infrastructure, and medical professionals willing to spend hours labeling images at the per-pixel level.

Mango Medical has been building this since 2020. The company was founded in Freising, Germany, near Munich, which sits in one of Europe's denser orthopedic clinical networks. That proximity to academic medical centers and implant company R&D labs matters. GDPR compliance in the EU adds complexity but also gives the team regulatory familiarity that transfers to FDA work.

The eight-figure letter of intent from a major orthopedic company that the team has disclosed suggests at least one of the big players has concluded the training data and clinical validation are credible enough to negotiate a real partnership. That is not a trivial signal at this stage.

The Landscape: How It Stacks Up

Surgical and orthopedic AI is a crowded category in our data at StartupHub. Across the 270 surgical and orthopedic AI companies we track, the average total score sits around 34. Mango Medical scores 32 today, which puts it squarely in early-commercial territory. The leaders in adjacent areas, Viz.ai (74) and Aidoc (70), have years of regulatory clearances and commercial deployments behind them. Augmedics (69) takes a hardware-augmented-reality approach to spine surgery, a different technical bet entirely.

The key distinction: most surgical AI companies are building visualization tools, navigation aids, or outcome prediction models. Mango Medical is building decision automation for the preoperative planning workflow. That is a different regulatory category, a different buyer, and a different integration surface. Closer peers are companies like Blueprint Medical (acquired into Stryker) and Zebra Medical's planning tools, which targeted similar workflows before being absorbed into the implant giants. The consolidation trend in this space is a feature, not a bug, for a startup angling for acquisition or deep commercial partnership.

Difficulty Score

Building this is genuinely hard across every dimension:

ML and AI (9/10): Foundation models for 3D medical image segmentation, agentic planning pipelines, implant-specific constraint solving. The U-NET expertise is table stakes; everything above it is still cutting-edge research.
Data (8/10): Annotated CT scan datasets with IRB clearance, HIPAA/GDPR compliance, and ground-truth surgical outcomes are among the hardest training datasets to acquire in any industry.
Backend (7/10): DICOM ingestion pipelines, HIPAA-compliant cloud infrastructure, API design for enterprise healthcare IT integration, clinical validation data management.
Frontend (4/10): Surgical plan viewer, annotation interfaces, and report generation are comparatively standard; the hard part is not the UI.
DevOps (6/10): Medical AI deployment under FDA Software as a Medical Device guidelines, audit logging, model versioning with clinical traceability, HIPAA Business Associate Agreements with every cloud vendor.

The Moat and What Is Hard to Replicate

Three things are genuinely hard to copy here:

First, the training data. Four years of annotated clinical imaging data, assembled with patient consent and regulatory compliance across EU and US healthcare systems, cannot be recreated quickly or cheaply. A competitor starting today would need three to five years and significant capital just to reach parity on dataset quality.

Second, the founding team's medical imaging expertise. Jorge Padilla Perez's background is not a marketing credential. It means the architecture choices are right, the labeling tooling is designed correctly, and the clinical validation methodology is sound from the start. Recruiting a comparably credentialed team is possible but expensive and slow.

Third, FDA regulatory progress is a legitimate time-based moat. A 510(k) application takes twelve to eighteen months minimum. Once cleared, the predicate status itself becomes a competitive asset: future competitors have to demonstrate equivalence to Mango Medical's cleared device, not to some older, less capable predicate.

What is easy to replicate: the API wrapper, the web frontend, and the general agentic architecture. A well-funded team could replicate the software layer in six months. They just would not have the data or the clearance to deploy it in a clinical context.

Replicability Score: 78/100

Deep proprietary training data, active FDA clearance process, U-NET-pedigreed founding team, and an existing eight-figure commercial relationship put Mango Medical firmly in high-moat territory. The thing that keeps this from a 90 is that the implant industry is highly consolidated, the big players have acquisition budgets, and the moat changes character entirely if Stryker or Zimmer Biomet decide to acquire rather than partner. At this stage, the most likely outcome is not independent growth to unicorn scale. It is a high-multiple exit to an orthopedic giant that wants to vertically integrate AI planning and stop paying third-party engineering firms to do it manually.

For founders looking to build in adjacent spaces: the wedge of replacing expensive human expert workflows inside established medical device sales processes is underexplored. The implant company's planning engineer is a $150,000 salary problem that occurs at every sale. Automating that with validated AI is a measurable ROI story that enterprise procurement teams can approve without needing a C-suite champion.

Mango Medical picked the right starting point, assembled the right team, and now has a regulatory clock running. If they close FDA clearance before a well-funded competitor gets into the segment, the exit optionality from that position is substantial.

© 2026 StartupHub.ai. All rights reserved. Do not enter, scrape, copy, reproduce, or republish this article in whole or in part. Use as input to AI training, fine-tuning, retrieval-augmented generation, or any machine-learning system is prohibited without written license. Substantially-similar derivative works will be pursued to the fullest extent of applicable copyright, database, and computer-misuse laws. See our terms.

Build This Startup with Claude Code

Complete replication guide — install as a slash command or rules file

# Build a Mango Medical Clone: 7-Step Developer Guide

## Step 1: Data Infrastructure and DICOM Pipeline

Set up a HIPAA-compliant data environment before writing a single model line. Use AWS GovCloud or Azure Government for PHI storage with Business Associate Agreements in place. Build a DICOM ingestion pipeline using pydicom or dcm4chee to ingest, validate, and de-identify CT scan files on arrival. Schema: studies table (study_uid, patient_id_hash, modality, acquisition_date, body_part), series table (series_uid, study_uid, slice_thickness, pixel_spacing), and instances table (sop_uid, series_uid, s3_path, pixel_data_hash). Store raw DICOMs in S3 with versioning and immutable audit logs in CloudTrail. Every access to PHI must be logged with user_id, timestamp, and action.

## Step 2: Medical Image Segmentation Model

Implement a U-NET variant for 3D bone segmentation. Use nnU-Net (the self-configuring medical image segmentation framework) as your baseline. Install: pip install nnunetv2. Prepare training data in nnU-Net format: each case is a folder with CT volume (0000.nii.gz) and segmentation mask (seg.nii.gz). Label classes: background (0), cortical bone (1), cancellous bone (2), cartilage (3), pathological region (4). Train with: nnUNetv2_train DATASET_ID 3d_fullres FOLD. For shoulder anatomy, you need minimum 200 annotated cases for reliable segmentation. Use ITK-SNAP or 3D Slicer for annotation tooling. Target Dice score above 0.92 on held-out test set before moving forward.

## Step 3: 3D Anatomy Reconstruction and Implant Sizing

After segmentation, extract surface meshes using marching cubes (scikit-image or VTK). Align each patient scan to a canonical atlas using rigid registration (SimpleITK). Compute anatomical landmarks: glenoid center, glenoid version angle, humeral head center, canal diameter. Map these measurements to a component lookup table encoding sizing rules for each supported implant system. Build the component database as a PostgreSQL JSONB table: implant_systems (id, manufacturer, product_line, component_type, sizes, measurement_constraints, contraindications). The sizing algorithm is a rule-based constraint solver over measurement outputs; start with decision trees, move to ML only if rules fail on edge cases.

## Step 4: Agentic Planning Pipeline

Build the orchestration layer using a multi-step agent. Define tools: segment_anatomy(dicom_path) -> segmentation_mask, reconstruct_3d(mask) -> mesh_and_landmarks, size_implant(landmarks, implant_system) -> component_recommendation, validate_plan(recommendation, anatomy) -> validation_result, generate_report(recommendation, validation) -> structured_plan. Use the Anthropic API with tool use for the orchestration layer. The agent calls tools sequentially, runs validation checks (canal diameter clearance, glenoid bone stock sufficiency, range-of-motion simulation), and iterates if a check fails. Hard-code fallback: if confidence < 0.85 on any step, flag for human review rather than returning a plan.

## Step 5: API Design and Enterprise Integration

Build a REST API with endpoints: POST /v1/studies (submit DICOM, returns study_id), GET /v1/studies/{id}/status (polling endpoint), GET /v1/studies/{id}/plan (returns structured surgical plan JSON), POST /v1/studies/{id}/feedback (surgeon accepts/rejects/modifies plan). Plan schema: { study_id, patient_ref, procedure_type, implant_system, components: [{type, size, laterality, placement_params}], approach, flags: [{severity, message}], confidence, generated_at }. Authenticate via API keys scoped to implant company accounts. Rate limit per account. Log all plan retrievals for audit. Use async job processing (Celery + Redis) since CT analysis takes 30-120 seconds.

## Step 6: Clinical Validation and Regulatory Documentation

For FDA 510(k), you need: predicate device identification (search FDA 510(k) database for cleared surgical planning software), substantial equivalence argument, software documentation per IEC 62304 (software lifecycle processes), risk management file per ISO 14971, clinical performance study comparing AI plans vs. expert engineer plans on a blinded hold-out set of 50 or more cases. Metrics: component size match rate (target above 95%), plan acceptance rate by surgeons, revision rate. Document everything in a Design History File. Use Greenlight Guru or Qualio for QMS software. Plan 12-18 months from submission to clearance.

## Step 7: Deployment and MLOps for Medical AI

Deploy on HIPAA-compliant Kubernetes (EKS with encrypted volumes). Implement model versioning with clinical metadata: each model version gets a validation_study_id linking to the performance data justifying deployment. Never hot-swap models in production; deploy new versions to a shadow environment, run parallel evaluation for 30 days, then promote with change control documentation. Set up monitoring: track prediction confidence distributions, flag distribution shift. Use LIME or SHAP for per-case explainability logs, required for clinical trust and regulatory audit. Maintain a model card for each deployed version documenting training data demographics, known failure modes, and performance by subgroup (age, pathology grade, BMI).

Install for:

claude-code-skills.md

#YC W2026 #MedTech #Surgical AI #Orthopedic #Medical Imaging #Foundation Models #Deep Tech #Healthcare AI #DICOM #FDA