Netivei Ayalon · Transport Model QA
Beer Sheva Model Validation
How I built an AI-native system to reconstruct, validate, and package a complete transport model output — and ship it as a production tool to the customer.
Follow the storyThe question
From Raw Model Output to a Validated, Packaged Deliverable
Transport model outputs are complex — dozens of tables, thousands of rows, multiple time periods. Manually checking them against targets takes weeks and is error-prone. I built a system that does it automatically, and ships as a ready-to-use tool.
- My contribution
- Building independent table reconstruction and comparison tools for model-output review.
- The deliverable
- Reproducible calculations, comparison reports and an interface for tracing discrepancies.
- What it enables
- Making differences between reported and reconstructed results visible and reviewable.
The approach / 01
The system reads raw model outputs — road network assignments, transit ridership, route data — and reconstructs each required report table using the exact formulas from the model's specification document.
The approach / 02
Each reconstructed table is automatically compared against the official target. The system flags every mismatch — missing rows, extra rows, value differences above tolerance — and produces a detailed diff report.
The approach / 03
A professional Streamlit dashboard lets an analyst review any table, switch between time periods (AM, off-peak, PM), filter by match status, and inspect individual cell-level differences with smart color coding.
The approach / 04
The entire system was packaged as a standalone deliverable and handed to the customer — including all inputs, scripts, GUI, and documentation — so they can run it independently at any time.
The evidence
The System in Action
A walkthrough of the reconstruction dashboard and validation interface.
What you are seeingModel-output review. A flagged difference starts an investigation; it does not by itself establish which result is correct.
The detail behind a reported number.
Reconstructed vehicle-hours from table 6.1. At report precision, two values can look identical. Reveal more digits to inspect the difference.
- Reported
- 1,707.17 vehicle-hours · AM
- Reconstructed
- 1,707.17 vehicle-hours · AM
- Absolute difference
- 0.00 vehicle-hours · AM
The approach
Built AI-Native
The reconstruction engine, test suite, GUI, and packaging workflow were all built using AI-native development with Claude and Codex. Not just code generation — full-cycle development: architecture design, iterative implementation, automated testing, and production hardening. This is the workflow I teach.
Why It Matters
From evidence to a decision.
Independent reconstruction makes model outputs easier to question and verify. The system exposes missing rows and numerical differences, with calculation methods and source inputs available for review. Flagged differences remain visible until their cause is understood.
You can't approve a model you can't verify. This system makes verification fast, repeatable, and transparent.Get in Touch
The next perspective
BRT Corridor Analysis
Agent-based Pink Line corridor simulation, powered by Replan.
Explore case studyAll projects