portfolio - anselm long

TL;DR

Hack & Roll 2026 was chaotic in the best way possible.

We built Optifiner — an evolutionary, multi-agent AI code optimizer that:

spawns parallel agents to propose code changes
runs benchmarks/tests to score those changes
only keeps improvements that measurably beat the baseline

Website: https://optifiner.ai

Context

I came into this hackathon with the usual fear:

I’m about to embarrass myself next to people who ship faster than I can open VS Code.

Turns out building with people way smarter than you is… kind of amazing.

Not because they do everything for you, but because you get dragged into a higher bar:

sharper product decisions
cleaner engineering taste
less “let’s build a platform” and more “ship one good demo path”

What Is Optifiner?

Optifiner is basically a Darwinian loop for code.

Instead of “one agent edits files and hopes for the best”, it works like:

Spawn multiple agents in parallel (analyzer / refactorer / optimizer / etc)
Each agent proposes changes in an isolated workspace
Run an evaluator (benchmark + test gate) to produce a score
Keep only the changes that improve the score
Repeat across generations until it converges

The part I liked most is that it forces you to define what “better” means.

Because without a metric, “AI optimizer” becomes: random refactors with vibes.

What We Shipped

Basing this off our repo write-up (which is way more put-together than my sleep schedule), we shipped:

a multi-agent evolution loop (parallel runs)
benchmark-driven acceptance (improve score or revert)
git-integrated history (every improvement is a commit)
a dashboard to watch evolution progress
multi-model support (Claude / GPT / Gemini)

The Hackathon Experience

Best Welfare I’ve Seen At A Hackathon

I’m used to hackathons that are like:

“Here’s a banana.”
“Good luck.”
“Sleep on concrete if you want.”

Hack & Roll 2026 was the opposite. Legit the best hackathon I’ve been to in terms of welfare.

And yes, it changes the entire experience. Suffering builds character, but so does… not suffering.

We Won A Tiny Prize (And It Leaks)

We won a small prize from Marshall Wace — a Stanley bottle.

It leaks sadly. But it’s the thought that counts.

Building With Cracked People

Biggest reflection:

Building with people way smarter than you is a cheat code.

You end up asking better questions:

Is this the simplest thing that works?
How do we make the demo robust?
What’s the metric? What’s the failure mode?

Reflections

Benchmarks > vibes. If you can’t measure it, you’re doing interpretive dance.
Parallelism is your friend. Letting agents explore different approaches feels like the AI version of splitting tasks.
Shipping a demo is a separate skill. “It works locally” is not a demo.

Thanks

Huge thanks to my teammates: Philippe, Verity, and Leroy.

What’s Next

We’re going to keep working on this after the hackathon too (because apparently we enjoy pain).

Hack & Roll 2026: Building Optifiner, a Multi-Agent, Evolutionary Framework