Bonsai demo · recorded run · 2026-11-11

Compliance in a box.
Bonsai on one GPU.

A folder of 103 certificate files goes in. A compliance dashboard, a CSV report and 6 draft reminder emails come out, in 46.9 seconds. Ordinary code reads the files. Bonsai 2 27B, running on one local RTX 5090, is the brain: it plans the run, reviews the checks, answers questions and drafts the emails.

46.9 sfolder in to artifacts written
8 of 8planted problems found
10Bonsai calls, all served by Bonsai 2 27B
127.0.0.1the only address the pipeline connected to

What happens

A messy folder in. A dashboard out.

The data is synthetic: 103 invented certificate files (94 PDF, 9 PNG, 15 of them scans with no text layer) for a fictional firm, Northgate Civil Engineering, with 41 invented names. Eight problems were planted in the folder on purpose.

1 · Extract

Ordinary code, no AI. A certificate extractor, called as a library, pulls 101 certifications for 41 names in 0.06 s and folds 2 older renewals into current licences.

2 · Check

Code checks flag problems. Expiring this month, expired, a cert number shared by two people, a near-duplicate name, a missing expiry.

3 · Think

Bonsai is the brain. It plans the run, reviews every flag and explains it in one plain sentence, answers two questions, and drafts the reminder emails.

Read this first
  • Not an offline run. The network was not blocked. The claim is narrower: every connection the pipeline made went to localhost, and the model was pinned to the local GPU.
  • No OCR. The 15 scans were read from their filenames only. The dashboard marks them low confidence.
  • Extraction is ordinary code; Bonsai is the brain. The model does not read the documents. It plans, reviews, answers and writes.

The full run · 94 s

Watch the whole recording

Uncut screen recording of the run: the folder, the agent log as Bonsai works, then the dashboard it produced. No sound.

The output

The dashboard it built

This is the dashboard file the run wrote, unchanged. It is self-contained and uses synthetic data for a fictional firm. Click through Alerts, Ask Bonsai and Email drafts.

Run sheet · run id 1d2bd679f40f

How it was measured

Every number below comes from the recorded run's own metrics file. Nothing here is estimated.

Numbers measured in the recorded run

Wall time, folder in to artifacts written46.9 s
Documents in103 (94 PDF, 9 PNG); 15 scans with no text layer
Certifications extracted (code)101 for 41 names, in 0.06 s; 2 older renewals folded into current licences
Status65 active, 3 expiring, 2 expired, 31 with no expiry (30 OSHA cards that print none, plus the planted one)
Bonsai calls10, all served_by: bonsai2-27b; 0 cross-model fallbacks; 0 cache replays; thinking off
Bonsai tokens1,886 out; 9,232 in as reported by the server (the review call reported 0 prompt tokens, so the true input is higher)
Bonsai speed41.5 tok/s per call (median), 15.7-66.2 range; 41.3 tok/s overall. These are end-to-end per call, including prompt processing and queue wait, not pure decode speed
GPUNVIDIA GeForce RTX 5090, all layers on the GPU; utilisation 98% peak, 55.6% mean; 22,139 MiB peak
Shared loadThe Bonsai server was also serving other traffic. Sampled every 2 s: 1-3 of its 4 slots busy (ours used at most 1)
Network, pipeline process1 TCP connection, to 127.0.0.1:8150 (the local model scheduler). DNS: 127.0.0.1 only. Logged by a socket.connect audit hook
Network, browser199 requests, all to 127.0.0.1:18946; 0 non-local requests (anything else would have been blocked)
RoutingEach call was pinned to the local backend with local_only: true, so the queue fails closed rather than routing to another model

Planted problems: 8 planted, 8 found

Who found itResult
Code checks8 of 8 flagged (3 expiring this month, 2 expired, 1 shared cert number, 1 near-duplicate name, 1 missing expiry), 0 false positives
Bonsai reviewConfirmed all 8 and explained each in one plain sentence
Bonsai: "Who expires this month?"3 of 3 correct, nothing extra
Bonsai: "Anything suspicious?" (no hints about the planted problems)Named the shared cert number and both expired certs, plus 2 of the 3 that expire this month. Missed the near-duplicate name and the missing expiry. Added one true point that was not planted: two NICET records have no number (they are scans read from the filename)

Honest caveats

  • Not an offline run. The network was not blocked. The claim is narrower: every connection the pipeline made went to localhost, and the model was pinned to the local GPU.
  • No OCR on this box. The 15 scans were read from the filename only, and the dashboard marks them low confidence.
  • Extraction is deterministic code, not AI. Bonsai 2 27B is the agent brain: it plans the run, reviews the code's checks, answers two questions and drafts the emails.
  • Two earlier takes were thrown away: in one the clock froze on screen, in the other there was a 29 s queue wait. Before routing was pinned, one dry-run call was answered by another model; it is pinned now, and the run fails if served_by is anything other than bonsai2-27b.
  • Dates are relative to 2026-11-11. The synthetic folder is regenerated for whatever day the run happens.