I’ll tell you what I got wrong about mentorship before I tell you what I got right. It’s more useful that way.
When I started mentoring at DeepChem — the open-source deep learning library for drug discovery — I made the mistake every technical person makes. I was too prescriptive. I knew the codebase. I knew the architecture. I knew where the tricky parts were. So I’d lay out exactly how to approach a problem, which functions to modify, what the test suite expected. The mentee would execute the plan competently and learn almost nothing.
The plan was the problem. When you hand someone a map, they follow it. When you hand them a compass, they learn to navigate.
The GSoC Model
Google Summer of Code pairs open-source projects with student contributors for a summer of real work — not toy problems, not tutorials, but production code that ships. Over multiple GSoC cycles at DeepChem, I worked with more than a dozen PhD and postdoc fellows on projects ranging from neural ODEs to transformer models for genomics to conformal prediction frameworks.
The project that crystallized everything was the 2022 D-MPNN implementation. Aryan Amit Barsainyan came in as a GSoC contributor and delivered 18 merged pull requests and over 4,500 code additions in a single summer. The final implementation was competitive with the original Chemprop codebase — a serious molecular property prediction model, not a weekend hack.
But again — that’s not the story. The story is what happened between the pull requests.
What Actually Works
Somewhere around my fifth or sixth mentee, I stopped planning their work and started planning their environment. The shift was subtle but load-bearing:
Set the problem, not the solution. “We need a D-MPNN implementation that integrates with DeepChem’s model API and passes these benchmark tests.” Full stop. How you get there is yours. I’ll be available when you get stuck, and you will get stuck, and that’s the interesting part.
Make the codebase legible, not the task. Instead of explaining the task, I’d spend the first week helping the mentee read the existing code. Once they understood the architecture — really understood it, not just the function they needed to modify — the task became self-evident. They could see what was missing because they could see what was there.
Normalize the struggle. I’d tell every mentee: “The first two weeks will feel like you’re drowning. That’s correct. You are drowning. It gets better around week three when you realize the water isn’t as deep as you thought.” Naming the difficulty honestly — without minimizing it or dramatizing it — gives people permission to be confused without feeling like confusion means they don’t belong.
Celebrate the process, not the person. Not “you’re so talented” — that’s fixed-mindset praise that makes people afraid to fail. Instead: “That debugging strategy was smart. Walk me through how you figured out the issue was in the message-passing layer.” Praise the approach. Make the reasoning visible. Let them see their own growth as a skill, not a trait.
The Pattern That Kept Repeating
Aryan’s trajectory wasn’t unique. It was the pattern. Contributor arrives uncertain. First PR is small and careful. Something clicks around week three — not confidence exactly, but fluency. They start reading code they weren’t assigned, asking questions about architecture instead of implementation, proposing changes I hadn’t considered. By the end of the summer, they’re not executing my vision of the project. They’re executing theirs.
That’s what I got wrong early. I thought mentorship was knowledge transfer — I know this, now you know it too. It’s not. It’s creating conditions for someone to discover what they’re capable of. The mentor’s job is to make themselves unnecessary as fast as possible.
Why This Matters Beyond Code
The DeepChem mentorship pattern — real problems, real tools, insistence on impact, compass not map — is the seed of everything I’ve built since. The World Genome Academy puts DNA sequencers in high school classrooms using the same philosophy: trust students with real tools and real problems, provide structure without prescription, measure outcomes by what they produce, not what they memorize.
Science Sundays at The KINN works the same way. Show up with a question, not a credential. Pod leaders don’t lecture. They create the conditions for unexpected connections.
The deep knowledge needed to build an AI agent is in a human mind. The deep knowledge needed to do open-source drug discovery is in a community. Mentorship is how you make that knowledge accessible without flattening it. Not “let me simplify this for you.” More like: “Let me stand next to you while you figure this out. I’ll tell you when you’re close.”
Bharath Ramsundar built DeepChem on the conviction that open-source drug discovery could democratize deep learning for science. What I learned from working alongside him and mentoring in that community is that the democratization happens person by person, pull request by pull request, confused question by confused question.
I can’t help but believe as a stupid optimist that this is how science actually advances. Not through genius. Through community, through patience, and through the willingness to hand someone a compass and trust them to find their own way.