Docflow Labs

The first AI native BPO in healthcare. Outsource to AI for cheaper and more reliable workflows.

[Jonathan Milgrom @jbmilgrom]

we (@docflowlabs) shipped something similar internally although for a heavier usecase - automating browser based workflows from a session recording - so its a bit heavier

also an electron app with an embedded browser but uses rrweb for higher fidelity dom signal, can track multitab, comes with a playback viewer, and can ship with a native package for transcription of the audio. the electron app session produces an artifact that a sister "session cli" is designed to understand (with a corresponding skill of course for agent digestion)

maybe we should opensource

[Kyle Jeong @kylejeong]

https://t.co/ZRdHQha7y3

[Docflow Labs @docflowlabs]

boosh

[Jonathan Milgrom @jbmilgrom]

@kylejeong We've doing something similar at @docflowlabs
(see link below)... placing AI in a buildtime loop iterating on an artifact, as opposed to at runtime (e.g. stagehand action cache).

Does stagehand fit in here / is it going to evolve in that direction?

https://t.co/d8P42duTIB

[Jonathan Milgrom @jbmilgrom]

@_ScottCondron @altryne @karpathy We’ve also been doing this for a bit at @docflowlabs

https://t.co/qnBMEJUjk0

[Ilan Buckman @IBuckman]

@toddsaunders We do something like this at @docflowlabs. Nice to see it becoming a pattern.

[Docflow Labs @docflowlabs]

Software is still mostly all you need

It's neural nets as buildtime accelerators and traditional software at runtime still, together giving use reinforcement learning with software as the substrate

From our CTO @jbmilgrom

https://t.co/QYXnfIayTo

[Ilan Buckman @IBuckman]

People often jump straight to “AI agents,” but providers can't rely on something non deterministic. Ask an LLM the same question ten times and you'll get different answers.

What providers actually need is a HIPAA compliant system that follows clear, repeatable logic. If BCBS sends code A, bill code A. If it fails, try B... Humans get bogged down doing that branching logic over and over. Computers can do it all day and all night.

The interesting part is how AI makes this accessible now. Coding these workflows manually used to take huge engineering budgets, so only large healthcare companies could afford real automation. Now AI can build these deterministic workflows cheaply and handle the “human only” parts too, like reading documents that look different each time.

At @docflowlabs we use AI to build payer specific workflows for DME providers and let them run in a deterministic, HIPAA compliant way. Feel free to DM.

[Docflow Labs @docflowlabs]

Quick PSA on the claude code 2.0 update. It appears that the "think" keyword is no longer recognized.

Instead, there is a new "tab" hot key (not shift + tab, which rotates between regular, accept edits and planning modes) that rotates between thinking on and off. We were under the impression for a full day that sonnet 4.5 was 10x faster and perhaps a little worse than Opus 4.1, when actually we just accidentally had thinking off.

With thinking on, our impression is that sonnet 4.5 is on par or maybe a little better than Opus 4.1 and more like 2 or 3x faster.

[Docflow Labs @docflowlabs]

@jbmilgrom https://t.co/VVrInSiysm

[Docflow Labs @docflowlabs]

buildtime agents > runtime agents by our cofounder @jbmilgrom

[Docflow Labs @docflowlabs]

At docflow, every job is an opportunity for automation

[Jonathan Milgrom @jbmilgrom]

claude code rn 😂... bro you _are_ a script