Mackenzie is the Global Startup Evangelist at AWS. His days are spent traveling the globe to meet startups, share their stories, and connect engineering teams together. Every day there are a large number of startups launching on AWS across every imaginable industry. It’s Mackenzie’s mission to find stories of startups that are helping to improve the world and share these stories with a wide audience.
AWS AI League: Forge Your Own Specialist Model
A model customization challenge. This season: Drug Discovery.
There are more AI models available today than any team can keep up with, and the biggest, most expensive one isn't always the one your problem needs. Come fine-tune a small model into a custom specialist that outperforms frontier, general-purpose models on your task and runs at a fraction of the cost. You'll evaluate models, select the right one, optimize it through prompt engineering, then push it further by fine-tuning it on real data.
This season, we run it through one of the highest-stakes problems there is: drug discovery. Your model will take on one of two challenges that mirror how real drug candidates get judged: integrate multi-modal data to identify a disease-specific therapeutic target, or predict a molecule's properties like permeability straight from its SMILES string, the chemical shorthand for its structure. These are the same questions that decide whether a candidate makes it to the clinic, and the same approach Insilico Medicine used to build TargetPro and Nimbus Therapeutics used to build their Novus model on AWS.
The domain is drug discovery, but the skill is universal: evaluate, select, and customize a model to win at your business's specific problem. Whatever you build here, you'll leave knowing how to do it for your own.
You can team up with others to prepare training data, configure fine-tuning jobs on SageMaker AI, and submit predictions to an automated evaluation pipeline that scores against experimental measurements in real time. Sample datasets like TargetBench and ChEMBL in JSONL Q&A pairs will be provided, and you can bring your own datasets as well. The team with the most accurate predictions wins up to $3K in AWS credits.
Hands-on build sprint: use Amazon Quick to prepare datasets for model fine tuning; take advantage of serverless model fine tuning capability on Amazon SageMaker AI to develop your customized model. Deploy your best model to Amazon Bedrock. Use Amazon Quick Desktop as your scientific workbench to interact with your own model for competitive intelligent analysis.
Karan is a technical AI product leader with 12+ years of experience progressing from software engineering through applied science to product management. He currently owns product strategy and P&L for AWS's LLM customization products, partnering with enterprise customers and internal businesses on AI strategy, including evaluation design and the accuracy, latency, and cost tradeoffs that determine deployment success.
With over 5 years at Insilico Medicine, Petrina's work reflects a commitment to leveraging cutting-edge AI platforms to address complex challenges in the lifesciences. Guided by a mission to drive transformative advancements, I emphasize collaboration and innovation to enable groundbreaking discoveries and bring next-generation therapies to market more efficiently.
Â
A fine-tuned model you built, hands-on experience with Amazon SageMaker AI, and potential prizes up to $5K in AWS credits.
Prizes:
Category: Target Identification
1st Place: $3,000 in AWS Promotional Credits + AWS AI League Championship invitation
2nd Place: $2,000 in AWS Promotional Credits
3rd Place: $500 in AWS Promotional Credits
b. Category: Molecular Property Prediction
1st Place: $3,000 in AWS Promotional Credits + AWS AI League Championship invitation
2nd Place: $2,000 in AWS Promotional Credits
3rd Place: $500 in AWS Promotional Credits
T&C's are available here.Â
Â
