01 / The Problem

Most cancer patients never get a tumour board.

A multidisciplinary tumour board — where pathologists, researchers, and oncologists deliberate together — is the gold standard for complex cancer cases. But they take weeks to convene, require specialist co-location, and are unavailable in most of the world.

2–4 weeks

Average time to convene a tumour board

<20%

Of cancer patients globally access one

35%

Of diagnoses changed after board review

02 / Three Agents Enter the Room

A digital board that reasons, debates, and decides.

GigaPath Pathologist

Researcher

Lead Oncologist

Five specialist AI models collaborate sequentially — vision, language-vision, retrieval, staging, and synthesis. Each model runs simultaneously in the MI300X's 192 GB unified memory.

03 / The Pathologist Sees Pixels

GigaPath Pathologist

GigaPath sees what the human eye misses.

Prov-GigaPath (1.1B parameters, trained on 1.3 billion pathology tokens) encodes every 224×224 patch into a rich morphological embedding. Attention rollout reveals which regions drive the diagnosis.

🔬 Nuclear atypia detected in 3 of 12 patches

📊 MC Dropout uncertainty: 91% ± 4.2%

🧬 EGFR mutation morphology pattern: high probability

04 / The Researcher Cites NCCN

Every recommendation is cited.

RAG over pre-indexed NCCN guidelines, TCGA studies, and PubMed abstracts. The researcher retrieves evidence, reranks it, and challenges the oncologist if the initial plan misses a guideline.

NCCN NSCLC Guidelines 2024

Category 1

FLAURA Trial — Osimertinib

Phase III RCT

EGFR Mutation Testing Recommendations

CAP/IASLC

Researcher

05 / The Debate — Agents Change Each Other's Minds

The researcher caught a missing EGFR test.

📚 RESEARCHER CHALLENGE

⚠️ EGFR status unknown. NCCN Category 1 for first-line TKI only applies to EGFR-mutant. Recommend molecular testing before committing to chemotherapy.

👨‍⚕️ ONCOLOGIST REVISION

Acknowledged. Revising first-line pending EGFR result. Adding osimertinib as conditional Category 1 recommendation.

−+Revision diff — what changed after debate

First-line: Osimertinib 80 mg/day (EGFR-mutant, NCCN Category 1, FLAURA trial). Platinum-based chemotherapy (cisplatinif +EGFR pemetrexed).negative.

Consensus Score: 87/100This is what distinguishes AOB from a single-LLM answer.

06 / The Math Behind the Architecture

This architecture is impossible on a single H100.

Toggle components to see the memory arithmetic. The MI300X's 192 GB HBM3 unified pool is not a performance advantage — it's what makes this architecture physically possible.

🧮 Interactive VRAM Simulator

Toggle components to see which GPU can hold the full AOB stack. Required components are locked.

AMD MI300X · 192 GB HBM3109 / 192 GB

192 GB

NVIDIA H100 · 80 GB limit💥 OOM +29 GB

← 80 GB OOM threshold

H100 cannot run this configuration. Requires 109 GB — 29GB over the H100's 80 GB limit. The MI300X has 83 GB headroom remaining.

07 / The Benchmark Proves It

82.3% TNM accuracy. 74.8% biomarker F1. Reproducible.

Evaluated on 100 curated clinical cases. All metrics with 95% bootstrap CIs. Dataset published on HuggingFace for independent verification.

Model / Config	TNM Acc.	Biomarker F1	Tx Alignment	Schema Valid.
AOBAOB Full Pipeline	82.3% [80.5%, 84.4%]	74.8% [72.9%, 76.9%]	77.8% [76.1%, 79.5%]	97.0% [96.0%, 98.0%]
No Debate Rounds	77.1% [75.2%, 79.1%] ▼ 5.2%	70.1% [68.0%, 72.2%]	73.4% [71.5%, 75.3%]	96.0% [94.8%, 97.2%]
No LoRA Specialists	74.4% [72.3%, 76.5%] ▼ 7.9%	67.8% [65.6%, 70.0%]	71.1% [69.1%, 73.1%]	95.0% [93.7%, 96.3%]
Single LLM Baseline	69.1% [66.8%, 71.4%] ▼ 13.2%	61.2% [58.9%, 63.5%]	64.4% [62.2%, 66.6%]	92.0% [90.4%, 93.6%]

95% bootstrap CIs (N=100 cases, 1 000 resamples). TNM Acc. = exact TNM stage match. Biomarker F1 = macro-averaged F1 over EGFR/ALK/ROS1/KRAS. Tx Alignment = NCCN guideline alignment. Schema Valid. = structured JSON output conformance.

Full Benchmark →Run a Case →