01 · Our Models

From static paper to predictive safety

Human-led safety verification cannot scale to today's complex aerospace architectures. Ellie rebuilds the fault tree from the paper; Mia works out every combination of events that fails it and where the hidden faults sit — before they can ever manifest at the aircraft level.

The Build Order

One foundation, three agents

Mia and the Investigator are the agents industry values most — and both are only possible once Ellie exists. So Ellie came first, Mia now runs on her output, and the Investigator follows.

01

Built — running on pilot trees

Ellie

Foundation agent · ingestion & structure

Ellie is the agent we built first, and the one every other agent depends on. She reads legacy fault-tree drawings page by page — vector PDFs today, scanned drawings next — and rebuilds the Boolean structure: gates, events, transfer symbols and the connections between them. Every page passes structural checks before an analyst sees it, and her review workflow points people at the pages that need a human rather than at every page.

Symbol and gate detection with a pose model trained on fault-tree drawing grammar, not generic OCR
Vector-geometry parsing that recovers gate fan-in exactly; transfers resolved across pages into one tree
Structural checks on every tree: common-cause exposure, circular references, orphan pages, dangling transfers
Annotation and repair in place — a flagged page is fixed where it is and re-checked
Measured: 99.6% node-level connection accuracy on held-out benchmark pages; 150–1,500 page packages

02

Built — running on Ellie's output

Mia

Analysis agent · quantification & latent faults · built on Ellie

Mia takes the tree Ellie compiles and answers the question the paper never could: which combinations of events fail the system, and where are the faults nobody wrote down? She quantifies the tree with a proven fault-tree engine, then hunts for what a document review misses and ranks the pages a reviewer should read first.

Exact minimal cut sets, path sets and importance measures (Birnbaum, Fussell-Vesely, RAW, RRW)
Latent-fault classes: single points, near-singles, exposed common cause, undeveloped and conditional events
Beta-factor common-cause groups, Monte Carlo uncertainty, pass/fail against ARP4761 probability classes
Review ranking per page — decisive, careful, skim — and a fleet-wide event library that surfaces inconsistencies
A grounded assistant that answers only from the analysed trees, on a local model; every answer traceable
Measured: the complete cut-set solution of a 150-page, 485-gate tree in 0.2 seconds

03

Planned

Investigator Agent

Root-cause agent · built on Ellie and Mia

The Investigator agent will trace incidents backward through the compiled failure chains and Mia's cut sets to isolate root cause — turning the same executable models into a forensic tool for fleets in service.

Trustworthy Framework

Built for zero-hallucination logic

Unlike generalized, off-the-shelf AI models that guess or hallucinate, our specialized multi-agent pipeline operates natively and exclusively on rigid fault tree data structures.

PHASE 01 · Ellie

Spatial parsing

Extracts precise Boolean gate mechanics (AND, OR, INHIBIT, k-of-n voting), transfer symbols, and primary event nodes directly from technical schematics.

PHASE 02 · Ellie

Relationship disambiguation

Connects the dots across pages and disjointed files — repeated event IDs become one event, transfers resolve into one tree — to spot structural anomalies no human eye could catch.

PHASE 03 · Mia

Continuous logical cross-check

Mathematically evaluates the compiled network: exact cut sets, importance, probability against ARP4761 targets, and the latent single points and common-cause exposures the paper never showed.