Hello! I am Aakash Tripathi, an AI research scientist working in precision oncology.

Portrait of Aakash Tripathi
Multimodal AI for Precision Oncology: From Data Integration to CDS, Stanford MedAI Group Exchange, session 155
Multimodal AI for Precision Oncology: From Data Integration to CDS · Stanford MedAI Group Exchange, session 155, 2026 (Image: Stanford MedAI video thumbnail)
Line illustration of a compact NAS case with hot-swap drive bays, network and power cables, and a speech bubble reading 2x parity and 32 TB online
The server: Homelab, part 1: the server, its storage, and its power
Multi-panel diagram of the HONeYBEE framework. A multimodal dataset store feeds four processing rows: histopathology (WSI loading, stain processing, tissue detection, patches), molecular (protein expression, DNA methylation, gene expression, miRNA expression, DNA mutation, then normalization and denoising), clinical (reports and EHR extracted into tokenized text chunks) and radiology (3D volumes sliced into patches). Each row ends in its own encoder, and the resulting embeddings go to an embedding dataset with public hosting. A side panel lists downstream tasks: cancer classification, survival prediction and patient retrieval, plus future treatment recommendation and biomarker discovery.
HONeYBEE: foundation model embeddings for multimodal oncology data. Reproduced from Tripathi et al. (2025), npj Digital Medicine, CC BY-NC-ND 4.0.
Accelerate Cancer Research With AI-Driven Multimodal Data Integration, NVIDIA GTC 2025, Poster P74176
Accelerate Cancer Research With AI-Driven Multimodal Data Integration · NVIDIA GTC 2025, Poster P74176, March 20, 2025 (Image: the poster PDF from the GTC session page, titled "HoneyBee: Developing Multimodal AI-Ready Datasets from Public Cancer Repositories")
Line illustration of a compact NAS with a stack of containers rising from it, next to app tiles for photos, media, documents and cameras
Self-hosted apps: Homelab, part 2: what I self-host
Three-column diagram of the extraction framework. Stage 1, extraction with justification: TCGA and Moffitt pathology reports plus a prompt go to 11 extraction LLMs (Qwen, Tulu3, Phi4, three DeepSeek-R1 sizes, Llama 3.1, 3.2 and 3.3, Mixtral, Mistral) that return site, laterality, histology, stage, grade and behavior with justifications. Stage 2, automated evaluation: the reports, the extractions and an evaluator prompt go to three reasoning LLMs (DeepSeek-R1 large, Qwen, QwQ) that assign a score to each LLM. Stage 3, aggregation via reasoning: DeepSeek-R1 large produces the final output with explanations, followed by expert validation. Below each stage is a sample output for a bladder urothelial carcinoma report.
CLEVER & pathology reports: Consensus reasoning with local LLMs for pathology report extraction. Reproduced from Tripathi et al. (2026), Laboratory Investigation, © 2025 USCAP, published by Elsevier; reused under Elsevier author rights.
Grid of 64 H&E tissue patches extracted by HONeYBEE
Scalable Multimodal AI in Oncology Using HONeYBEE: From Embeddings to Clinical Impact · Mayo Clinic AI Summit, July 8, 2025 (Image: HONeYBEE patch-extraction output, lab-rasool/HoneyBee, CC BY-NC-ND 4.0)
Block diagram: imaging data (RadImageNet, 1000 dims), clinical data (demographics, labs, staging and markers, 10-15 dims) and text reports (radiology reports and pathology notes, 1024 dims) each feed an encoder of linear, BatchNorm plus ReLU, linear, dropout and linear layers that outputs 128, 32 and 128 dimensions. The three encoded vectors enter a cross-modal attention mechanism, then a feature fusion network with two outputs: a primary Cox risk score for survival prediction and an auxiliary binary event prediction of death or censoring. A lower panel lists three attribution methods: attention attribution, gradient attribution and integrated gradients.
EAGLE: attention fusion and attribution for cancer survival. Reproduced from Tripathi et al. (2025), arXiv, CC BY-NC-ND 4.0.
Architecture diagram in four panels. A: sources (NCI Genomic Data Commons, with UCSC Xena, cBioPortal and future databases as expansions) send clinical and biospecimen metadata as TSV and JSON files. B: AWS Lake Formation with an S3 ingest bucket, AWS Lambda, AWS Glue and the Glue Data Catalog. C: AWS data warehouse (RDS MySQL, Redshift) and analytics (Athena, QuickSight). D: CRDC repositories PDC, IDC and GDC feed unstructured data download tools through their APIs, producing a data folder of EHR, raw data and a JSON manifest for the user.
MINDS: linking public cancer repositories into ML-ready cohorts. Reproduced from Tripathi et al. (2024), Sensors, CC BY 4.0.
Whole-slide image thumbnail beside artifact, background, and tissue probability maps from a DenseNet121 tissue detector
Tissue-detector output from our lab's HONeYBEE-based hackathon materials for the 2025 Dr. Robert Gillies Machine Learning Workshop in Cancer
Overview figure of the SeNMo self-normalizing multi-omics framework
SeNMo: learning pan-cancer prognosis from sparse multi-omics data. Reproduced from Waqas et al. (2024), arXiv, CC BY 4.0.
Four-panel schematic. A: a 101 bp DNA sequence is one-hot encoded into a 4 by L grid. B: a per-factor ConvNet expert maps the one-hot input to a 32-dimensional hidden layer, a fully connected layer and P(bound). C: the shared one-hot input feeds frozen ConvNet, DeepSEA, DanQ and DNABERT-6 experts for 7 factors, each producing a 32-dimensional embedding that enters a softmax gating network, then a weighted combine and fully connected layer that outputs P(bound). D: ShiftSmooth turns a noisy vanilla-gradient sequence logo into a smoothed one by shifting the input, running the MoE, shifting back and averaging.
HetMoE: Gating, not averaging: HetMoE for transcription factor binding sites. Reproduced from Tripathi et al. (2026), Mathematics, CC BY 4.0.

Recent news

  1. Started as Generative and Multimodal AI Research Scientist, Precision Oncology, at Moffitt Cancer Center.
  2. HetMoE and ShiftSmooth for transcription factor binding site prediction published in Mathematics.
  3. Speaker and co-presenter, Learning Lab LL4022 at the SIIM Annual Meeting in Pittsburgh: running local LLMs behind institutional firewalls.
  4. Two abstracts at the AACR Annual Meeting 2026 in San Diego: real-world multimodal AI and multi-agent extraction of social determinants of health.
  5. The Pathologist covered our pathology report extraction study.
  6. USCAP 2026 abstract on a multi-agent system for neuro-oncology biomarker extraction (San Antonio).
  7. Consensus-based LLM extraction from surgical pathology reports published online in Laboratory Investigation.

From the blog

All posts