Worldwide Studios 100 天 AI for Science 奖学金项目演示回顾：Benchmate、自动化细胞培养实验室等五个项目

You’re getting this note as you joined a recent event. If these aren’t for you, that’s cool too. You can unsubscribe below.

Hi Everyone,

Thanks for joining for our demo night! There were so many great discussions over where new machine learning and robotic workflows fit in the lab.

If you want to follow up with a specific fellow and what they built, more info below. We also launched applications for this fall’s free, virtual fellowship if you want to build new projects in community.

These demos are from a 100 day fellowship

These demos concluded a free 100 day fellowship, where each fellow built and shared something new every 2 weeks. The fellowship was hosted by Worldwide Studios, an education nonprofit helping people of all backgrounds build with the latest in AI and robotics.

Shout-out to the community mentors who joined for this one to support the projects: Luis, Di, Mo, Charvi, Haran, Simon, Srusti, Elvan, Nikhil, and more.

Learn more about the fellowship

DEMOS + RECAP

1. Natalie demo’d Benchmate, an AI labmate to close the loop

Benchmate is a free, open-source AI “labmate” from Natalie that walks through all four stages of the scientific process including pulling PubMed evidence and Geneformer perturbation predictions, generating and ranking hypotheses via a seven-agent system inspired by DeepMind’s Co-scientist, designing wet-lab protocols (including reading freezer boxes with CryoVision), and analyzing your raw results to propose sharper follow-ups.

It’s live at benchmate.streamlit.app with code on GitHub, and bring your own API key.

More info

• GitHub

• Blog post

• Connect more with Natalie here on LinkedIn.

2. Dale demo’d a DIY self driving cell culture lab, featuring the latest shaker

Dale, a mechanical CAD designer from industrial automation, is building a modular self-driving lab for cell culture viability studies that aims for benchtop reliability at scale, with his keystone build being a heater-shaker prototyped for roughly $100 in components using maker space electronics, salvaged printer stepper motors, and 3D printed PETG, controllable via local touchscreen, web client, or integration into an automated workflow. He also worked on a decapper with an industrial collet design, a flask tipper, a mobile robot, a salvaged six-axis arm, a gantry microscope, and a Blockly interface for an OT1, using Claude to calibrate hardware from photographs. Next up: finishing the instruments with food coloring tests, making them washdown compatible, and growing actual cells, and he’s looking for fellowship or after-hours collaborators.

More info

• GitHub

• Connect more with Dale here on LinkedIn

3. Sameer shared a computer vision pipeline with wearables for data driven improvements for playing tennis

Sameer, a lifelong tennis player, built a prototype to answer whether a bad performance is fatigue or technique, pairing wearable physiological data (readiness, sleep, HRV) with pose data extracted from his own match video via Google MediaPipe and custom Python to measure things like peak racket hand speed. His N-of-one findings showed a strong correlation between high readiness and higher swing speed, a wedge he says no one has connected before, since existing tools measure either physiology or tennis stats but not the link. He arrived here after pivoting from a reaction-time hardware device to a software version to a coaching dashboard.

More info

• GitHub

• Blog Post

• Connect more with Sameer here on LinkedIn.

4. Julia built an observability layer for agent runs using tools for experiment design

Julia built “watch agent run,” a repo that turns opaque agent logs into deterministic readouts (input/output tables, color-coded picture views, and a provenance reporter) so scientists can check what an agent claims it did against what the append-only logs show it actually did. Demoed on extracting macrophage differentiation protocols from papers, where all three prompting approaches looked identical at 8 papers but diverged sharply at 200, with every run claiming full coverage and only one actually achieving it. Her thesis: what you can see, you can begin to trust, and she’s asking life sciences researchers what they wish they could inspect.

More info

• GitHub

• Connect more with Julia here on LinkedIn

5. Vivek demo’d AI drug discovery engine for rare diseases

Vivek, a 20-year Genentech drug discovery veteran, built a rare disease platform that uses precomputed multi-level biological representations to power two world models, one generating personalized mechanistic disease hypotheses and one designing patient-specific AAV therapeutics, with Pareto frontiers threading the tension between biologist intuition and ML exploration, demoed on Duchenne muscular dystrophy. He built it because he cares about making a difference in a heartbreaking disease landscape, and credited the fellowship for the confidence that comes from actually making something.

More info

• GitHub

• Connect more with Vivek here on LinkedIn

UPCOMING FELLOWSHIP

This was our first 100 day fellowship for those curious to explore what’s possible in new scientific workflows.

We’re getting ready to launch our second one. It’s free to participate, virtual, and part time. Ship new projects in community. 8-10hrs a week, with 1x a week sync. Demo every 2 weeks.

First round of applications due Aug 28.

Learn more and apply here

Thank you to our venue host!

​Edison Scientific is building an AI platform that reads, reasons, and experiments across scientific disciplines from literature synthesis to hypothesis generation to full research workflows.

​A commercial spinout of the AI research lab FutureHouse, Edison is behind Kosmos, an autonomous AI scientist used by tens of thousands of researchers at universities, national labs, and biopharma companies.

Shout-out to Leina for all the help.

See you at a future community event,

Michael

注意转载授权 · 阅读原文