I would build tools that are helpful for wet lab scientist to deal with their pain points in their day to day work. Most times they don't know how we can help them, and we can be most helpful with some minor work. Some tools that would be helpful:

• Annotation of genes in pathways, ontologies, diseases and other resources: Most resources like GO, reactome, WikiPathways are heavily biased, scientist don't know how to contribute to those resources, and we only use them without contributing back. Having an easy way to contribute would help all, but this would require time to review which would exceed the 72h.
• Automate or help with tedious tasks they have: I had a researcher as for the IDs from a sequencing facility they provided because she needed it for another form but she couldn't copy from their websites (and there were 59 samples with long names). Ad-hoc websites for researchers to compare internal results (not yet on GEO or other public resources). Recently they spent quite some time designing a flow cytometry panel when perhaps a tool to do that would help them. These small tasks are the ones that hinder most (academic) researchers.
• How to design experiments preventing batch effects. Most tools are designed to account and correct for after the experiment is done, we could prevent them before if they had tools. A tool that helped prevent it would be helpful (if used ;)

I would build a platformed aimed at extracting models from distributed datasets without requiring data or metadata to be transferred over the platform; i.e., it operates local data and local compute, and only communicates API commands, parameter updates, or bulk statistics (which can be suitably perturbed to protect privacy). A user could:

1. Register data + metadata with the platform, sharing it to be used as training or validation
2. Identify registered datasets of interest (e.g., fMRI data with a neurological disease phenotype)
3. Request a uniform pre-processing of identified datasets
4. Fit a statistical model leveraging all datasets
5. Evaluate the model on the local dataset (or any datasets registered as "Public"); and interrogate the resulting model

Basically: Data silos definitely exist; but you don't have to de-silo all the data to benefit from "all" of it -- just enable the results of pre-specified privacy-preserving (in a zero-knowledge sense) computations to be shared. Throw in a tit-for-tat mechanism (the more you share, the more computations you can run) and you might have a way to encourage a weird sort of cooperation.

Edit: But first I would invent a way to fit a year's worth of time into a single day.