←
Brian Sam-Bodden5,391 posts
Brian Sam-Bodden avatar

Brian Sam-Bodden

@bsbodden

AI/ML Chief Scientist @Java_Champions Data Science @Harvard 🟧🟦

◷ Joined August 2007▧ 98 Photos and videos
Brian Sam-Bodden avatar
Brian Sam-Bodden@bsbodden·Oct 28

Four local models and three hosted providers, one router, rules instead of if-statements. 7 of 10 requests never left my laptop and the bill came to a third. Four of my own quality claims failed measurement — fixing them made the fleet better, not smaller. medium.com/@bsbodden_68805/m…

Brian Sam-Bodden avatar
Brian Sam-Bodden@bsbodden·Oct 7

Ribbon filters match Bloom's accuracy in ~30% less memory but cost ~4x the CPU to build. Preprint: one counting pass that groups keys into cache-sized windows makes Ribbon construction ~2x faster at scale. It halves the gap to Bloom. doi.org/10.5281/zenodo.23222…

Brian Sam-Bodden avatar
Brian Sam-Bodden@bsbodden·Sep 23

“Slow Down” good for the industry? Even if the reasons are 1) exaggerated claims for tech that might have reached its engineering asymptote, 2) Pre-nationwide-IPO for the new “oil” industry, right when the realization of “Hey, that toaster in the kitchen can also do fast matmul!”

James Ward avatar
⟳ Brian Sam-Bodden retweeted
James Ward@JamesWard·Sep 11

I get the reasons why people prototype in Python and eventually graduate to Rust. I go another way: prototype & scale with JVM languages. So I can move fast and get great runtime performance, without a rewrite in between.

Brian Sam-Bodden avatar
Brian Sam-Bodden@bsbodden·Aug 25

Java/Kotlin... I needed a way to see if agentic generated codebases were rapidly accumulating tech debt. So I wrote a little library that aggregates mature static code analysis tools into a geometric mean formula, returning a value between 0 and 1. MFCQI integrallis.github.io/mfcqi-…

Brian Sam-Bodden avatar
Brian Sam-Bodden@bsbodden·Jul 16

arxiv.org/pdf/2607.09691 digs into how much context an LLM needs once you located the code to work on. The edited code region carries most of the useful signal and providing the rest of the file is unnecessary in single-shot scenarios. Roughly 70% fewer input-context tokens!

Brian Sam-Bodden avatar
Brian Sam-Bodden@bsbodden·Jun 17

I’ve been experimenting with #GOAP in the context of Agents, mainly with Embabel. Since I also have a lot of projects on #LangGraph, I wanted to try to GOAPify some of them, so I put together a GOAP framework extension for LangGraph, aptly named LangGOAP github.com/LangGOAP/LangGOAP

Rowan Trollope avatar
⟳ Brian Sam-Bodden retweeted
Rowan Trollope@rowantrollope·May 5

Cool to see how OpenAI uses Redis to scale voice AI to 900M users: openai.com/index/delivering-… We're huge Codex fans here at Redis - I use Codex 5.5 extensively - amazing work by the team there. GPT 5.5 is the best model for most tasks IMO. It even fixed my wifi tonight (no joke)