Peyton Li

Data, SWE, ML | 1x Hackathon Winner | Data Science + Statistics @ UC Berkeley

[HACKERSQUAD @HackerSquadSF]

Really cool to see @peyton_li_ back building again at the @runpod Flash Hack Day.

He's working on an agent that fine-tunes small models for specific tasks, like writing SQL queries or building apps in Rust. The goal is to match the performance of much larger models.

Love seeing builders focus on practical, specialized AI tools.

More stories coming soon 👇
@HackerSquadSF #brightdata

[Zhihu Frontier @ZhihuFrontier]

🎉 10K followers! Thank you!
To celebrate, we're teaming up with @MiniMax_AI for a giveaway!
✅ Follow @ZhihuFrontier & @MiniMax_AI
🔄 Repost this post

🎁 We'll randomly select 10 winners to receive:
• 1-month MiniMax Token Plan (Plus level) with full access to the MiniMax model family (M3 / M2.7 / image / speech / music)
• 1-month Zhihu Knowledge VIP
📅 Ends July 4.

Every follow, every repost, every AI discussion has helped @ZhihuFrontier grow into the community it is today. ❤️
Here's to the next 10K—and many more conversations about AI. 🚀
#Zhihu #MiniMax #AI #Tech

[Peyton Li @peyton_li_]

Repo: https://t.co/KE6PU6128h
Research: https://t.co/UJEdfFhTfr

Thanks to @TheBuildersCollective @HackerSquadSF for hosting such an amazing event.

#Hackathon #AI #Agents #MachineLearning

[Peyton Li @peyton_li_]

I won 1st place at the AI Inference Hack Day. 🥇

The problem:

AI agents are slow. Every tool call, "pull the schema", "search vectors", "load history," is a cold 800ms roundtrip. A 10-step workflow takes 8 seconds. Now imagine a two or three hour grinding session.

[Peyton Li @peyton_li_]

Next turn, the /resolve endpoint measures actual cache hit/miss latency for a real reward and updates the model weights. Bandit converges in ~20 steps to 80-90% latency savings over naive caching.

Tech Stack: TypeScript, Next.js, React, Tailwind, Akamai, Gemini, Turborepo.

[Peyton Li @peyton_li_]

How it works: every agent turn produces a 12-dim context vector. The bandit picks a pre-fetch action using closed-form linear algebra. No gradient descent, no GPU, <1.8ms per predict. The winning arm triggers a pre-fetch.

[Peyton Li @peyton_li_]

So I built VeloxEdge, a predictive edge cache that figures out what the agent will need before it asks. It's a LinUCB contextual bandit that watches the agent's latent-space trajectory and pre-fetches assets to the nearest edge node.

[Peyton Li @peyton_li_]

@mynameisyahia @getcontextdev Building an app for college students to plan courses with AI. It auto adjusts when a course is unavailable and can also be adjusted based on number of credits you want per semester. I would use https://t.co/6ChTxWy16O to scrape all the data for major and college requirements.