Selected work

Case study, construction inspection

One report, two operating models.

A regional construction firm produced inspection reports the way most firms do: photos on phones, notes in a notebook, hours at a desk turning memory into prose. We rebuilt the workflow around a structured field-capture app, a database of every prior report, and a panel of specialist review agents coordinated by a QA orchestrator. Same deliverable. Different machine.

Before Single author, single review

The desk-bound report

Knowledge lives in the inspector's head and a shared folder of PDFs.

  1. 1

    Field capture, by hand

    Photos on a personal phone, observations scrawled in a notebook. Nothing is tagged, nothing leaves the device.

  2. 2

    Return to office

    Report writing waits until the inspector is back at a desk, often days later. Memory is already softening.

  3. 3

    Transcribe from memory

    Photos sideloaded, notes retyped, captions written from recollection. Detail loss is invisible until a dispute.

  4. 4

    Hunt for precedent

    Browse a shared drive of past PDFs by filename. Copy-paste boilerplate. Hope you picked a clean version.

  5. 5

    Single-author draft

    One inspector writes the whole report. Tone, math, and citations are only as consistent as one person on one day.

  6. 6

    Manager review

    One reviewer, finite attention. Catches some math errors, some contradictions, some missing sections. Misses the rest.

  7. 7

    Deliver and file

    PDF emailed to the client and dropped in a folder. The findings are now unsearchable.

    Output: PDF only
After Structured capture, multi-agent review

The instrumented report

Every observation is structured at the source and every draft faces a panel of reviewers.

  1. 1

    Field capture, synced in real time

    Inspector opens the project on a phone or tablet. Photos and voice or text comments are tagged with project ID, location, timestamp, and inspection category as they are taken, and stream to a project-scoped cloud store. Offline-tolerant. Searchable the moment the inspector leaves the site.

  2. 2

    Draft assembled from project data and archive

    The report builder pulls every tagged photo, comment, and measurement for the project into the right section of the right template, with precedent language already wired in from a database of every prior report extracted into structured fields.

  3. 3

    Agentic QA process

    Five specialist agents run against the full draft in parallel, each owning one kind of check. A QA orchestrator agent merges their findings, deduplicates overlapping flags, ranks by severity, and resolves conflicts when two specialists disagree.

    Specialist agents, all five running at once

    Math & units Technical writing Contradictions Mandatory inclusions Exclusions & compliance
  4. 4

    Reviewer sign-off, with audit trail

    One prioritized punch list, with every flag traced back to the agent that raised it and the passage that triggered it. The reviewer accepts, overrides, or sends back each flag. Every decision is logged, building the evaluation signal for the next release.

  5. 5

    Deliver and archive, structured

    Client gets the PDF. The system gets the PDF plus the structured record, which feeds the next project's precedent search.

    Output: PDF + structured archive

Cycle time, site to client

5 to 7 days to under 24 hours

Review surface area

1 reviewer to 5 agents + QA + human

Searchable history

Folder of PDFs to Queryable database

Field detail loss

Memory-bound to Captured at source

Indicative figures for this engagement.

What we built

The infrastructure behind the second column.

"AI integration" is the visible layer. Most of the work is the plumbing underneath it: the structured data, the capture surface, and the evaluation harness that makes the agents trustworthy.

01

Historical report database

Extracted every prior PDF into structured fields. Project, site, findings, photos, citations, and signatures all become queryable rows instead of pages in a folder.

Data foundation
02

Field capture app

iOS and Android app for photos, voice notes, measurements, and tagged comments. Offline-tolerant. Auto-syncs by project ID with full provenance.

Capture surface
03

Cloud sync & retrieval

Real-time ingest, indexing, and retrieval that the report builder consumes. One source of truth for every project's evidence.

Pipeline
04

Five specialist agents

Math & units, technical writing, contradictions, mandatory inclusions, exclusions & compliance. Each has its own prompt, evaluation set, and pass criteria. Versioned and tested.

Review panel
05

QA orchestrator agent

Aggregates findings across the five specialists, deduplicates, ranks by severity, and produces one prioritized punch list with citations for the human reviewer.

Orchestration

Get in touch.

Send a short note about what you are working on. No pitch, no pressure. If it looks like a fit, we will set up a time to talk.

Email Joe