Jeffrey Bander MD

Medical Director, Network Development Mount Sinai Hospital Network Chief, Cardiology Mount Sinai Hospital West

[Jeffrey Bander MD @jeffBander]

Repo (Apache 2.0): https://t.co/dU4xZq5Lne

Same story at the high end: on one M5 Ultra, Tom's Hardware measured a model-specific engine decoding GLM 5.3 Flash about 3x faster than llama.cpp. https://t.co/cnROXBnnMo

[Jeffrey Bander MD @jeffBander]

A 26B model in ~2 GB of RAM on an 8 GB MacBook Air, at 5–6 tok/s.

TurboFieldfare keeps Gemma 4 26B-A4B's 1.35 GB core in memory and streams each token's experts from SSD.

On a Mac, the engine is the upgrade. Ask Brian @robotnurseio, our lead AI engineer for local compute.

[Jeffrey Bander MD @jeffBander]

A fix passed our test gate. Then it never reached main.
Brian caught it this weekend in asor, our local LLM engine for Apple Silicon: the tool_choice "none" guard was tested on a branch, not on main. He cherry-picked it back with a unit test.
Test the commit you ship.

[Jeffrey Bander MD @jeffBander]

@FLShockDoc @dr_coops I agree people should make their own ambient Dictation systems customize. It’s the way they want it. We do the same thing.

[Jeffrey Bander MD @jeffBander]

Today at #TCT2026 we built an app live on stage with the audience. They voted, wrote the features, an agent built it, and it was on their phones. Open source so you can run your own: https://t.co/0awJ7konoK @TCTConference @crfheart @MountSinaiNYC

[Jeffrey Bander MD @jeffBander]

Today at #TCT2026 we built an app live on stage with the audience. They voted, wrote the features, an agent built it, and it was on their phones. Open source so you can run your own: @TCTConference @crfheart @MountSinaiNYC

[Jeffrey Bander MD @jeffBander]

Autonomy for agents is something you earn with evidence, one gate at a time.

We'll keep sharing what we build and what breaks at MSW Lab AI.

[Jeffrey Bander MD @jeffBander]

Everyone wants an AI agent that can drive any app on your screen.

We're building one that can't, yet. On purpose.

Here's how we phase a screen-operating agent: guardrails first, autonomy last.

[Jeffrey Bander MD @jeffBander]

"Measured" means more than a pass rate. Our next eval gate includes negative tasks (prompt injection, ambiguity, forbidden actions, stale screens) that must pass 100%, plus do-nothing and random-click agents as baselines.

If a random agent scores well, your eval is broken.

[Jeffrey Bander MD @jeffBander]

Small local models are good workers and bad judges. A new paper: 0.91 on clean tasks, 0.00 when sources conflict. Our design: the model proposes, tested code decides and hands off to a human. https://t.co/3SAZP1gYkx

[Jeffrey Bander MD @jeffBander]

Agents that operate clinical screens need guardrails first: a default-deny app allowlist, a cap on typed characters, a wall-clock limit, and no PHI in approval printouts. Read-only before write. That is how we are phasing our screen agent at LabQ.

[Jeffrey Bander MD @jeffBander]

@AISuperDomain Will it work with Citrix

[Jeffrey Bander MD @jeffBander]

From Prompts to Goals: Why AI Coding Agents Are Moving to /goal Workflows https://t.co/HmiyceQWig

[Jeffrey Bander MD @jeffBander]

@Jcontreras75 @Jcontreras75 yes and this is just the beginning of a new line of research that we hope will lead to better outcomes in SNF

[Jeffrey Bander MD @jeffBander]

AI is finally cracking a 60-year-old problem in cardiology: pulling the heart’s magnetic signal out of hospital noise.
Great @WIRED piece on magnetocardiography — the work we’ve been doing at @MountSinaiNYC with @SandboxAQ.
Pivotal trial launching soon. Proud of this team.
https://t.co/8T7don8Mrj