Brian Feeny

Principal Architect @awscloud | AI/ML Staff @Cornell | @Harvard alum | Technologist | Veteran. Let’s make the world a better place and have fun doing it!

[Brian Feeny @brianfeeny]

Semantic caching: serve the nearest stored answer.

Measured on 40 prompt pairs:

Paraphrase, safe to serve: 0.836
Near-miss, different answer: 0.911

The dangerous ones are closer. I built the fix and still wouldn't run it.

https://t.co/hDP8sYbyl1

#AWS #DynamoDB #LLMOps

[Brian Feeny @brianfeeny]

Trained an LLM router on 110k labels. It scored 0.84 AUC.

Then I shuffled the labels within each task, destroying all real signal. It still scored 0.838.

It had learned to recognize the task, not the difficulty.

https://t.co/OMod3MINSG

#MachineLearning #LLMs #AWS

[Brian Feeny @brianfeeny]

Budgets, guardrails, routing, caching, metering — everything you want at LLM inference time has to fit in one AWS Lambda. AgentCore Gateway allows exactly one interceptor per phase.

Built it as a plugin chain and measured each one:

https://t.co/uNoacwlWUq

#AWS #Bedrock #LLMOps

[Brian Feeny @brianfeeny]

Your LLM gateway probably isn't your bottleneck.

Bedrock reserves max_tokens up front, and on Claude 3.7+ each output token burns 5 tokens of quota. Your gateway counts ~1:1 — so it sees headroom that isn't there.

Full architecture writeup:
https://t.co/82dXLy7Leg

[Brian Feeny @brianfeeny]

@airebathsUS @airebathsus I went to your Chicago location on September 19th. I left my bathing suit in the lobby. I called immediately but no one called me back.. I followed up and got no call back. I called on September 27th and got no call back. Please help me.

[Brian Feeny @brianfeeny]

@renato_mariotti Too bad he wasn't a politician, he could have just walked

[Brian Feeny @brianfeeny]

@AndyOstroy But look at the data when assault rifles were banned, those deaths dropped by 43%. Then when they were reinstated, those deaths rose by over 200%. It made a difference.

[Brian Feeny @brianfeeny]

@ethanschoonover I noticed that the https://t.co/yWL2pnR7Jl download from your site isn’t working

[Brian Feeny @brianfeeny]

@davenewworld_2 @CheriJacobus But why would it be “several” and not “all”. Why should any of these John Does be kept secret?

[Brian Feeny @brianfeeny]

This looks like its shaping up to be the best alternative to Twitter. https://t.co/jO8ybL8fpk

[Brian Feeny @brianfeeny]

@MuellerSheWrote

[Dr Mia de Kuijper @MiadeKuijper]

It was fascinating to attend Dean Emma Dench
@HarvardGSAS hosting #Truth and #Objectivity in the Era of #Fake News-a discussion between @nycscribe
@dsallentess @harvard and @mschudson2
@columbiajourn at the #HCNY with a large audience of alumni as part of GSAS-150 #Celebrations

[Brian Feeny @brianfeeny]

@votevets @natsechobbyist As a Vet, I agree!

[Brian Feeny @brianfeeny]

@AdamKinzinger @NormOrnstein So he paid $44B just to have Trump reject his offer to come back. What a mess

[Brian Feeny @brianfeeny]

@TheTNHoller @NormOrnstein @JuddLegum “Billions” is a lot of money, is there evidence of this?

[Andrea Chalupa @AndreaChalupa]

An important lesson of the @elonmusk tragedy is that people are not disposable

[Alex Parker @AlexParkerDC]

Sergey Brin revives Google Plus team from cryogenic chamber.

"It's time."

[Brian Feeny @brianfeeny]

@jtleek If the distribution of data is changing then its data drift

[Brian Feeny @brianfeeny]

@KariLake Your twitter profile is BS. Please update it to “EX-candidate for Arizona Governor”, but please leave the part about Trump endorsing you, receipts are important.

[Danilo Poccia @danilop]

Interesting walkthrough 👉 Generate images from text with the stable diffusion model on Amazon SageMaker JumpStart https://t.co/r2MNGnxag0 #AWS #MachineLearning