London AI x Science Hackathon Best use of Devin

Run a Parallel Lab.
Not one idea at a time.

Real research groups don't test one hypothesis at a time. Point five or more Devins at your track problem, one idea each. Every session ends in a PR with a measured result, and GPU work runs on Modal. You set the direction and judge the results.

Opens one Devin session that saves every playbook into your org, macros included.
Quickstart

Five minutes.
Then five experiments.

01

Add the playbooks.

Click "Add all 7 playbooks". One Devin session saves them all into your org. No copy-paste.

02

Add your secrets.

Add your Modal token under Settings → Secrets. Child sessions only see saved secrets, never ones pasted into chat.

03

Prepare the repo.

Run !hack_setup. It installs the sponsors' official agent skills, and every later session inherits them.

04

Run the lab.

Run !parallel_lab with your problem, metric and 5+ ideas. You get one PR per idea and a leaderboard.

Playbooks

One lab.
Every sponsor tool.

"Create in Devin" opens Devin's new-playbook page with the full text already filled in. Press Save. DeepMind and Anthropic tools are left out on purpose.

Credits & secrets

Save them once.
Every session inherits them.

Add these in Devin under Settings → Secrets, using these exact names. Only Modal is required for the challenge.

ToolSecret namesGet itCredit
Modal · requiredMODAL_TOKEN_ID
MODAL_TOKEN_SECRET
modal.com/settings/tokens$150 per builder
Hugging FaceHF_TOKEN (write)huggingface.co/settings/tokensCoupon for Jobs + ZeroGPU
Paperclip by GXLPAPERCLIP_API_KEYpaperclip.gxl.ai/keysFree, unlimited at the event
AmassAMASS_API_KEYplatform.amass.tech → API Keys$500 per team
The challenge

Best use of Devin.
Run a Parallel Lab.

Use Devin as your team's research engineers, while your team focuses on direction and judgment. !parallel_lab covers every requirement by default.

015+ sessions in parallel.

Each one tests a different idea: a model, a method, an optimisation or an analysis.

02A PR with a result.

Every session ends in a branch or PR, scored with the team's shared eval script.

03GPU work on Modal.

Each PR records the Modal app link, GPU type and cost.

04One leaderboard.

Deltas and uncertainty, with negative results kept in. That's your 2-minute demo.