Witify Logo Witify Logo

A whole house in photos. AI makes the inventory.

OrionFlow is TSA’s contents claims software, designed and built by Witify. AI classifies each photographed item; evaluators validate, price and hand over the claim file.

Client
Sector
Contents claims
Technologies
Laravel, Vue, Horizon, MySQL

01 The numbers

Six figures, all taken from the software itself

Six measures of the system itself: its cadence, its rules, its settings.

  • 30min

    between two AI classification passes A set schedule: at each pass, the AI picks up the items that are ready.

  • 5

    photos at most per item sent to the AI It returns a category, a name, a description and a confidence score.

  • 2

    confidence thresholds, set by the firm Below either one, a person decides.

  • 12

    item categories to start with The firm renames them, adds more, and sets the columns of each.

  • 10

    modules switched on or off, firm by firm From cleaning to replacement search.

  • 1,686

    automated tests Across the whole system.

Figures taken from OrionFlow’s code and screens on November 13, 2026.

02 Before

Every photo sorted, every item described by hand

After a fire or water damage, every item is photographed on site. Before OrionFlow, the photos were then filed by hand in shared folders, and each item was described, classified and valued in a large Excel workbook.

Fig. 01 The inventory workbook: one item per row, typed in by hand

03 The workflow

From photos to a locked file, in six screens

One file, followed from the photos of a room to the insured’s approval. The AI classifies; people group, validate, price and approve.

01 Group

Group a room’s photos into items.

Photos arrive sorted by room. The operator selects the ones that show the same item and creates an item, numbered within the file. The sub-folder’s name gives its location.
Fig. 02 Photo grouping, room by room

02 Go-ahead

Nothing goes to the AI until the operator says so.

Once grouping is done, the operator marks the file ready for classification. The processing pipeline then shows where the file stands, step by step, and what is waiting for a person.
Fig. 03 A file’s processing pipeline

03 Validate

Whatever the AI is unsure of waits for a person.

The AI suggests a category, a name and a description for each item, with a confidence score. Confident items land in their category’s tab; the others wait in the Human validation queue, where the evaluator validates several at once.
Fig. 04 A file’s Human validation queue

04 Price

A replacement price, then the depreciated value.

In an item’s record, the replacement search suggests equivalent products with their price and merchant. The evaluator links one to the item; the depreciated value is then worked out in the table.
Fig. 05 An item’s record and its replacement search

05 The table

Each category reads like a spreadsheet.

One tab per category, one item per row, the columns the firm chose. Cells are edited in place, and the totals follow.
Fig. 06 Loss assessments, Electronics tab

06 Approve

The insured approves, the adjuster locks.

In the client table, the insured approves each item or leaves a comment, and the evaluator follows the approvals row by row. Once the review is done, the adjuster or the administrator locks the file: it becomes read-only.
Fig. 07 Client table, approvals in progress

04 Thresholds

The AI suggests. Below the threshold, a person decides.

For each item, the AI returns a category and a confidence score. The firm sets two thresholds and writes its own classification rules, in plain sentences. Anything that falls below a threshold waits for a person.

Fig. 08 A firm’s classification rules and confidence thresholds

05 The file

One file carries the timeline and who sees what

Once the inventory is under way, the file is also where the claim’s steps are tracked and where the adjuster decides what the insured can see.

Timeline

Twelve possible steps, switched on file by file.

Each step carries its dates and its state; the file’s duration is counted in business days.

Client permissions

The adjuster chooses what the insured sees.

One switch per tab of the file, set file by file.

06 Firms

Built for one firm, ready for several

In the fall of 2026, OrionFlow was rebuilt to host firms other than TSA: each one gets its own address, data and settings. An administration centre opens a firm and decides which modules it uses.

Administration centre

Ten modules, ten switches.

A module switched off here does not show up for the firm.

Firm settings

Each category sets its own columns.

Drag and drop: the items table follows, in every file of the firm.

07 Our choices

Three choices to keep the AI a tool, not a risk

Three decisions mattered more than the rest. For each one: the risk to avoid, what we chose, and what TSA gains.

  1. A threshold decides what a person reviews.

    The risk
    An AI that gets it wrong does not say so. Reviewing everything cancels the gain; reviewing nothing lets the error reach the insurer.
    Our choice
    Every AI answer carries a confidence score. Below the threshold the firm sets, the item waits in a separate queue. An item classified by hand is never sent back to the AI.
    The gain
    Evaluators only review what deserves it, and keep the last word.

  2. Each item takes its own path.

    The risk
    An outside AI service has its limits and its outages. If every item in a file moves together, a single failure stops them all.
    Our choice
    Each item is processed on its own. A failure is retried later, five times at most, then shown with its reason.
    The gain
    A failure costs one item, and the file shows where it stands.

  3. Settings rather than code.

    The risk
    Software built around one firm’s habits: every new category, every new tax, every new firm becomes a development project.
    Our choice
    Categories, columns, taxes, steps, thresholds and classification rules are settings. Modules are switched on firm by firm, from the administration centre.
    The gain
    The firm adjusts its own tool, and a new firm starts from a template.

08 Under the hood

What OrionFlow is made of

The software’s building blocks, and what TSA put into service from February to October 2026.

Application
Laravel, Vue, TypeScript, MySQL database
AI
AI vision model, photo classification, reading a loss list in PDF
Replacement
Online product search, product, price, link and merchant
Files
The firm’s file storage, folders and photos of each claim
Processing
Horizon queues, background processing
Documents
PDF and Excel, in the file’s language, French or English
Quality
1,686 automated tests, across the whole system
Operations
Error tracking, backups, health checks

Key milestones

  1. February 2026 First real file Photos read from the firm’s file storage, AI classification, the assessment tables and the summary.
  2. April 2026 Adjusters and insureds Accounts for adjusters and insureds, timeline, PDF and Excel export, items to clean, file locking.
  3. June 2026 Photo grouping Manual grouping becomes the standard route. Electronics expertise and a new visual identity.
  4. August 2026 Budget and approval Budget reserve, item approval by the insured, to-do files, a language per file.
  5. October 2026 Several firms Categories, taxes and steps set by the firm, administration centre, a space for each firm.

Is your team still sorting photos by hand?

Thirty minutes to go through one of your files together.

Or write to us
Next case study Observatoire des tout-petits Read the case study

Names and figures shown in the interfaces are examples.