Keep approval outside the model
FastAPI persists the report and a durable job. LangGraph extracts observations and proposes matches, but the review service owns approval. The agent cannot silently turn a guess into official progress.
Agentic AI / Local prototype
Time Agent
A construction report becomes an evidence-linked proposal. A person checks the match before the official progress record changes.
01 / THE PROBLEM
Site reports describe work in natural language. Construction schedules use structured activities, units and versions. Connecting them requires preserving the source, resolving ambiguous activity names and separating a model’s suggestion from an authorized change.
02 / FOLLOW THE WORKFLOW
A simplified example of the documented flow. Sample outputs are authored illustrations, not live model predictions.
Source: sample site report
“Level 2: installed 12 window frames today.”
A supervisor reports completed work.
Extracted observation
12 window frames · Level 2
Keep the reported quantity attached to its source.
Pending approval — ledger unchanged
Activity: Level 2 window installation
A reviewer checks the proposed schedule match.
Illustrated approved update
+12 frames · source retained
Only an approved proposal becomes recorded progress.
Step 1 of 4 · Report
03 / ENGINEERING
FastAPI persists the report and a durable job. LangGraph extracts observations and proposes matches, but the review service owns approval. The agent cannot silently turn a guess into official progress.
PostgreSQL stores jobs and checkpoints. Leases, bounded retries and idempotent persistence support recovery without duplicating a proposal after an interrupted run.
Source locators and versioned schedules travel with observations. Corrections and retractions preserve history. Native schedule file imports are distinct from live vendor synchronization.
Approved progress can be exported as CSV or JSON. A transactional outbox handles retries and receipts to a mock PMIS adapter. This delivery path is explicitly a prototype.
04 / EVIDENCE & LIMITS
The 26 October 2026 role rehearsal covered 16 worker pairs and ranked 64 composed assignments. All 120 proposals passed validation; 118 met the declared targets. These are single-report component results, not 64 independent graph runs or a production accuracy score. The repository also separates deterministic software tests from live-model studies.
Production acceptance is incomplete. Representative reviewed data, a frozen assembled configuration, fresh validation and holdout, and sustained capacity checks remain open. Earlier assisted graph runs exceeded the 120-second readiness target. Scanned PDFs require reviewed transcription; image reports and automatic OCR are unsupported.
05 / READ THE SOURCE
Based on public repository documentation reviewed on 7 November 2026. Results are attributed to the project’s recorded evaluations; they were not rerun for this portfolio.
TOOLS BEHIND THE WORK
Explore the stack. Drag the ribbon or select a tool.