Flag · tracker vs reality
Project rated fully healthy in the tracker. The chat shows the client actively reassessing the engagement.
surfaced by cross source audit
The workshop · Operations intelligence
Software that reads the night shift.
For companies whose reality lives in chat groups and self reported trackers. This is the system I built and run in production, anonymized.
The problem
Service operations run on self reported trackers and thousand message chat groups. People rate their own projects healthy. Payments arrive and never get logged. Clients go silent and nobody counts the days. The truth is in the chats, and nobody has time to read them.
The system
A nightly pipeline reads every client and vendor conversation across the portfolio. Facts are extracted once, at ingestion, into plain database rows, so the morning dashboard opens instantly and never calls a model just to load a page.
A second model audits the first one's output against every source and raises what a human reviewer would miss. Every fact carries a link to the exact message it came from. And a project with zero client messages is forced to unknown sentiment; the system is not allowed to guess.
Inside the build
Project Lantern
Udaipur · 14 March
62
health
sentiment · guarded
collection · 38%
silence · 6 days
open asks · 4
ranked actions
Revised decor quote requested, unanswered 2 days
⌾ msg 412
Advance received, not logged in the tracker
⌾ msg 397
Confirm garland vendor by Friday
⌾ msg 388
compiled 03:16 · every fact cited
Tracker says healthy · chat says reassessing
high
Value unchanged after a scope cut
medium
2 flags · routed to a human
What it caught
Flag · finance
Payment received weeks ago, never logged as collected. The books said zero.
chat evidence linked · corrected same day
Flag · silence
Client silent 68 days with 11 decisions pending. A recovery play was prescribed in the morning brief.
computed from raw timestamps, not memory
Representative recreations · details redacted
87
projects, nightly
147
flags, first audit run
$14
a night, full sweep
0
model calls to load a page
The engagement
- Map the sources. Chats, trackers, sheets, wherever your operation actually lives.
- Benchmark models on your real cases, judged blind, so the cheapest model that passes earns each job.
- Ship the nightly loop with guardrails in code and a receipt on every claim.
- Add the audit layer, measure its precision honestly, and hand over the keys.