Corso Labs: ED Documentation Review Room

Hackathon participant buildIndependent submission - trademarks belong to their owners - synthetic demo only

MultimodalMulti-agentEnterprise workflow

Why this matters in emergency care

Fast inference turns an agent workflow from a waiting-room task into an interactive review.

The demo sends the same raster electronic medical record image and same physician-note review task through Cerebras and a GPU path. The point is practical: if a multi-agent chart review comes back in seconds, an ED clinician can inspect, edit, and move on while the encounter is still live.

Start here - click first

Live agent progress

Waiting to start: both providers will launch together from one server timestamp

0.00s until run starts

Intakesort facts

Imageread JPG

Workuporganize tests

HPIdraft story

Claim Checksupport claims

Local Rulesflag context

Safetycheck claims

Mergefinal review

Cerebras0.00s0.00s0.00s0.00s0.00s0.00s0.00s0.00s

GPU0.00s0.00s0.00s0.00s0.00s0.00s0.00s0.00s

Agent responsibilitieswhy each step exists

Triage, vitals, timelineBuilds the case timeline from arrival, symptoms, vitals, and clinician concerns.

Reads raster imageReads the non-selectable EMR image for ECG, troponin, CXR, and vital facts.

Groups tests/resultsGroups resulted and pending tests so the note does not bury key findings.

Drafts note storyTurns the raw encounter into clinician-editable documentation language.

Checks medical supportChecks whether note claims have medical-source support or need review.

Flags hospital contextFlags local assay, repeat-timing, and pathway details that vary by hospital.

Blocks overclaimsRemoves diagnosis, treatment, disposition, and validation overclaims.

Merges final packetMerges all checks into the final physician note review packet.

1 Input

Stitched electronic medical record image

raster JPG

Synthetic demo image generated for this project. Not selectable text; Gemma must read the pixels.

2 Cerebras output

Physician note review packet

Output appears here after you run the live side-by-side benchmark.

not run

HPI Draft

Waiting for run

Exam / Status Summary

Waiting for run

Result Synthesis

Waiting for run

Medical Decision Reasoning

Waiting for run

Image extraction

Key facts read from pixels

Vitals: not extracted yet

ECG: not extracted yet

Troponin: not extracted yet

CXR: not extracted yet

3 Speed benchmark

Benchmark waiting

press the red button to measure Cerebras against the GPU path

Cerebras0.0s

GPU0.0s

Wait reduction--not run yet

Relative speed--8-agent workflow

First token--not run yet

Output speed--not run yet

Claim support check

Shows whether each note claim came from the image, a medical reference, or a hospital-context flag, then marks what the clinician must verify.

Local-context flags

Troponin assay, repeat timing, and local pathway caveats.

Safety boundary

No diagnosis, treatment, disposition, or clinical validation claim.