David Mashiah is a supply chain and distribution expert who brings AI into logistics.
At a glance
- Name
- David Mashiah · דוד משיח · David Mashiach
- Field
- Supply chain & distribution · AI and data analysis
- Based in
- Tel Aviv, Israel
- Education
- Mechanical engineering undergraduate, Tel Aviv University
- Languages
- Hebrew · Arabic · English
- ORCID
- 0009-0004-4684-955X
- Entity record
- Wikidata Q140694415
- Open research
- 6 outputs, every one with a DOI
Overview
He worked hands-on in perishable-goods distribution: seven months in flower supply, buying direct from Israel's largest fresh-flower wholesaler and building bouquets to a costed spec, which took vase life from about half a week to 14–21 days and cost him sales anyway. Alongside it he builds complete systems (front end, back end, data layer, deployment and the tooling that keeps them maintainable) across domains that have little in common with one another: clinical monitoring, real-time multiplayer, and local-first desktop software.
Mashiah is based in Tel Aviv, where he is a mechanical engineering undergraduate at Tel Aviv University. The software is built alongside the degree rather than out of it.
What connects the work is not a subject area but a question: how do you know when a model has stopped telling you the truth? It is the same question in a distribution centre and in a metamaterial surrogate, and answering it is the through-line in everything he ships.
Method
The method is to build the thing and then measure where it breaks, rather than reporting where it works. MetaXFam is the clearest case: a surrogate model that scored almost perfectly on held-out data, and then, tested on unit-cell families it had never seen, fell below a baseline that looks at nothing except how much material is left. Twenty of thirty configurations did. That number is the finding, and publishing it was the point.
MetaXFam-D, published as MetaXfam22 and rebuilt under the new name in August, is the follow-up, and the discipline shows twice over. Across 231 ways of choosing which families to train on, the worst case moved by a factor of 390 — but none of the nine cheap rules for choosing well survived a change of model, including the one his own earlier paper had pointed at. Then an audit of his own data found that 61.6 per cent of its cells were duplicates, a rasterisation artefact that inflates in-distribution scores, and he republished the benchmark rebuilt around the finding rather than leaving it to be discovered.
The same discipline runs through the rest. The vernier demonstration on the home page reports a threshold in arcseconds and then derives why a display cannot reach the clinical range, so the words not for diagnostic use are a conclusion rather than a disclaimer. The error-correcting code work is checked against exact analytical results before it is checked against anything convenient. In a supply chain the equivalent question is which forecast to stop trusting, and when. That is a measurement, not an intuition.
Work
-
Research prototype
Qeben
A home-monitoring application for age-related macular degeneration, pairing a daily hyperacuity self-test with a triage workflow for the clinician receiving the alert. Developed with advisory input from ophthalmologists and university researchers acting in a personal capacity.
-
Live
1 vs 100
A television-format quiz game built for Tel Aviv University, in which one contestant faces a crowd joining from their phones. Each room is a single Cloudflare Durable Object, so its state is serialised by construction rather than by locking, and WebSocket Hibernation holds idle connections open unbilled.
-
Local-first
Leviathan
An application for study, training, habits, finance and planning, built as a static web application inside a desktop shell with no server component at all. SQLite natively on the desktop, SQLite compiled to WebAssembly in the browser.
Research
Open-access outputs, released under ORCID 0009-0004-4684-955X and indexed on Google Scholar. Independent work rather than peer-reviewed journal publication.
-
Journal contribution
Algebraic Encoding of Amino Acid Sequences
Extending finite-field models to GF(32).
doi.org/10.6084/m9.figshare.32885588 -
Software
GF(25) Encoding & Reed–Solomon Error Correction
A reproducible toolkit for error correction of synthetic biopolymer data, the same family of codes that repairs a scuffed barcode or QR label in a warehouse instead of only detecting the damage.
doi.org/10.6084/m9.figshare.32963456 -
Dataset
MetaXFam
A cross-family dataset and toolkit for mechanical metamaterial homogenization surrogates, measuring where machine-learning surrogates fail on unit-cell topologies they were never trained on.
doi.org/10.6084/m9.figshare.32993597 -
Preprint · under review
Perfect In-Distribution Accuracy Does Not Imply Learned Physics
The manuscript behind the benchmark: a surrogate can score perfectly on the families it was trained on and still have learned no transferable physics. Under review at Extreme Mechanics Letters.
doi.org/10.5281/zenodo.21534865 -
Dataset
MetaXFam-D
A cross-topology benchmark rebuilt to hold no duplicate geometries, after rasterisation collapse turned 61.6 per cent of the previous version’s cells into copies of one another.
doi.org/10.5281/zenodo.21597150 -
Software · with preprint
Sub-cell Eigenmode Descriptors
Averaging is a projection: it keeps the unit cell's two rigid translations and discards every other way the material inside can move. Version 2 recomputes the study on duplicate-free data, which removed version 1's central claim and left a smaller, steadier one in its place.
doi.org/10.5281/zenodo.22137052
He also writes an engineering newsletter, Engineering Frontier.
Constraints
Mashiah works primarily in Hebrew and Arabic alongside English, which means right-to-left layout is a starting condition rather than a later adaptation. Several of his projects target users who cannot comfortably read small type, which sets a specification most software never has to meet: WCAG AAA contrast, a 16-point floor on body text, 48-pixel minimum touch targets, colour never used as the only signal, and a night mode treated as a first-class theme rather than an inversion filter.
His personal site implements these same rules and exposes them as live controls, so a visitor can raise the type floor, switch every low-emphasis colour to AAA, disable motion, or mirror the entire layout into Hebrew and watch the page comply. It is also published in Hebrew in full, as דוד משיח.
An honest scoreboard
One of these systems reached a university stage, one reached clinicians and stopped, and one runs its author's own life. None has yet become a business. Mashiah states this plainly on his own site rather than presenting the work as a track record it is not; a position he takes because in software the reasons a project stalls are usually more instructive than the reasons another one shipped, and because a claim that cannot be checked is worth less than an admission that can.
Notes
The working notes are where the reasoning is shown rather than summarised, each one measuring where something breaks instead of reporting where it works. They are collected in the notes section, grouped by subject: supply chain, machine learning and measurement.