David Mashiah

Supply chain · Distribution · AI & data

David Mashiah

Also written David Mashiach · in Hebrew, דוד משיח

I am a supply-chain and distribution expert who brings AI and data analysis into logistics. I have run perishable distribution where a wrong forecast means spoilage, and I publish the research on why prediction models fail and how to keep data honest under noise. I build the systems these operations need, and I know when they are lying.

Supply chain & distribution AI & data analysis Research: ML · metamaterials · codes

Watch a forecast break → See the three systems

0
Model configurations that failed out of distribution
0
Perishable distribution, bought direct at wholesale
0
Research outputs with DOIs
$0
Infrastructure at 1,000 concurrent users

Every forecast is right until conditions move. I build the systems that know when they have stopped being right.

The differentiator, stated plainly

01 The chain

One chain, six handoffs, one clock

This is the object I work on. Goods move along the bottom, data moves along the top, and shelf life drains underneath both; it is the only one of the three that never refills.

  1. 01 Grown
  2. 02 Packed & labelled
  3. 03 Cold store
  4. 04 Distribution centre
  5. 05 On the road
  6. 06 On the shelf
  1. 01

    The forecast layer

    Every number on that chain is a prediction: how many, how soon, at what price. A model that scores perfectly on the conditions it was fitted to is not evidence that it works; it is evidence that nobody has asked it anything new yet. I measure where they invert, and I published the measurement.

    Watch one break →
  2. 02

    The identity layer

    A pallet is only what its label says it is. Reed–Solomon codes over finite fields, the family I published a working toolkit for, are why a scuffed barcode or a wet 2D label still resolves to the right SKU instead of quietly resolving to the wrong one. The RFID alongside it works differently: it detects a bad read and asks again.

    The published work →
  3. 03

    The clock layer

    Perishable stock does not wait for a better decision. I spent a year selling flowers into retail, which is where the arithmetic stops being abstract: being late is not a delay you recover from next week, it is the whole unit, written off, plus the shelf it was standing on.

    Where this comes from →
02 Method

How I know when a model is lying

A model that scores perfectly is not evidence that it works. It is usually evidence that you have not yet tested it on anything it has not already seen.

I trained a network to predict how stiff a sheet of metal is once you punch holes in it. It scored perfectly on the held-out test set. Then I changed the shape of the holes and it became worse than useless: worse than deleting the model and answering with the average of everything it had ever been trained on. That happened on 20 of 30 tests.

This is the failure that costs money. A demand forecast is a model fitted to conditions that already happened; the day those conditions shift (a new SKU, a disrupted lane, a holiday that refuses to behave like last year's), it does not degrade gracefully, it inverts. Most of the industry reports the version of the model that cannot fail. I measure the version that can.

So the work is verification, not only construction. In MetaXFam the engine is checked against exact analytical results before it is trusted to generate anything: a homogeneous cell reproduces its input stiffness to 1.5e-16, a rank-1 laminate matches exact laminate theory to 5e-15, and the auxetic physics falls out of the geometry alone. The same rule governs everything else here: nothing is reported as working until something has tried to show that it does not.

  1. 01

    Test outside the band it was fitted on

    A held-out score drawn from the same distribution measures memory, not generalisation. Change the conditions, then look.

  2. 02

    Compare against something trivial

    A model is only useful if it beats answering with the average. Twenty of thirty did not, and that is the finding.

  3. 03

    Validate against exact results first

    MetaXFam's engine reproduces a homogeneous cell to 1.5e-16 and exact laminate theory to 5e-15 before it is trusted to generate anything.

  4. 04

    Report the failure, not the fit

    Anyone can publish the configuration that worked. The number worth having is where it stopped working.

  5. 05

    A threshold is nothing without an instrument

    The demonstration on this page returns arcseconds, then derives why this display cannot reach the clinical range. That derivation is the result.

  6. 06

    State the floor

    One pixel at 60 cm subtends about ninety arcseconds. Every claim above it is bounded by that, and saying so costs nothing.

03 Constraints

The requirements I keep choosing

I am not a clinician. I pick problems that arrive with awkward, measurable requirements attached: perishability, deadlines, accessibility law, hard concurrency limits. These are the conditions that follow from them.

DomainRegulated health software · higher-education systems
Interface languagesHebrew (RTL) · Arabic (RTL) · English
Contrast targetWCAG AAA
Type floor16pt body · 24pt critical
Touch targets≥ 48px
Motion120–320ms · disabled under prefers-reduced-motion
ColourNever the sole cue, always paired with an icon and text
Single sourceEvery value declared once; a duplicate fails the build
Third-party requestsNone; every font and asset served from this domain
Peak concurrency1,000 players in a single room · measured, not estimated
Run cost at that scale$0
This pageImplements all of the above; the controls are in the corner
04 Work

Three systems, shipped end to end

One clinical, one real-time, one local-first. Each taken far enough that its constraints stopped being theoretical, which is the evidence, not the boast.

One is live for a university at a thousand players. One reached clinicians and stopped. One runs my own life. All three shipped. That is the claim.

The scoreboard, project by project

Qeben

Never deployed

Home monitoring for age-related macular degeneration. A daily hyperacuity self-test that looks for macular change between clinic visits, and the triage workflow that puts the resulting alert in front of someone who can act on it.

It did not reach clearance and was never deployed. It is here for the accessibility and architecture work, not as a product claim.

StatusResearch prototype · never deployed
Advisory inputOphthalmologists and university researchers, in a personal capacity — no institutional collaboration
Intended usersAdults 50+, losing central vision
InterfaceHebrew-first RTL · AAA · 16pt floor · 48px targets · night mode as a first-class theme
BackendMulti-tenant · one bounded context per module
RegulatoryEngineered to SaMD standards · not clinically validated · labelled not for diagnostic use

1 vs 100

Tel Aviv University

A live TV-format quiz game. One hero against a mob of up to a thousand, joining with a PIN and a nickname. The interesting part is not the game; it is holding a thousand live sockets in one consistent room without a race condition and without a bill.

Live · cup.taueng.com ↗

Peak concurrency1,000 players, single room
Run cost$0 · roughly 50 full games per day
Concurrency modelOne Cloudflare Durable Object per room: serialised by construction, not by locking
TransportWebSocket Hibernation · idle sockets held open unbilled
Rules enginePure, no I/O · 100% unit coverage
VerificationLoad-test harness simulating N players, so the number is measured

Leviathan

Local-first

A life-OS for study, training, habits, finance and planning. No server, no API routes, no server actions. One storage engine on both targets, so there is one set of behaviours to reason about instead of two that agree most of the time.

ShellTauri 2 desktop · Next.js static export
ServerNone
StorageSQLite native on desktop · SQLite-WASM on OPFS in the browser
Source of truthThe database; stores are in-memory caches over it
MigrationsHand-written SQL · Drizzle for queries and types
DiagnosticsSchema version, migration ledger, row counts, query timings, N+1 flags, one button
05 Research

Published work, with DOIs

Open outputs, released under ORCID 0009-0004-4684-955X: one journal contribution, two software releases, two datasets and one preprint. Every research claim on this page points at one of these, and the full record with abstracts is a page of its own.

  1. 01

    Algebraic Encoding of Amino Acid Sequences

    Extending finite-field models to GF(32). Journal contribution, 2 July 2026.
    doi.org/10.6084/m9.figshare.32885588

  2. 02

    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. Software, 11 July 2026.
    doi.org/10.6084/m9.figshare.32963456

  3. 03

    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. Dataset, 15 July 2026.
    doi.org/10.6084/m9.figshare.32993597

  4. 04

    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. Preprint · under review, 24 July 2026.
    doi.org/10.5281/zenodo.21534865

  5. 05

    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. Dataset, 29 August 2026.
    doi.org/10.5281/zenodo.21597150

  6. 06

    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. Software · with preprint, 31 August 2026.
    doi.org/10.5281/zenodo.22137052

06 Demonstration

A forecast that scores perfectly, then breaks

The failure my research measures, made touchable. Drag the conditions away from the ones the model was trained on, and watch the score fall past the point where predicting the average would have done better.

Demand model · least squares, fitted in your browser 60 training points · scored against the mean baseline

This is a demonstration, not a forecasting product. The demand curve is synthetic and the model is a real least-squares fit computed on this page; nothing is sent anywhere. It shows a mechanism, not a benchmark of any deployed system.

Step 1: The model is fitted

Sixty observations, every one of them drawn from a single narrow band of conditions: the shaded strip below. A linear model is fitted to them by least squares when this page loads. Inside that strip it is excellent, and that is the number that reaches the slide.

Step 2: Move the conditions

Now ask it about somewhere else: a price point the business has never run, a lane that reroutes, a supplier who behaves differently. The model has not changed, and neither has the underlying reality, only which part of it you are asking about.

R² where tested
·
Versus predicting the average
·
Mean error
·

Solid line: what demand actually does. Dashed line: what the model believes. Shaded strip: the only conditions it ever saw.

An R² below zero is the whole point: it means the model is doing worse than if it ignored every input and answered with the average. That is not a hypothetical. Across thirty train-on-one-family, test-on-another combinations in MetaXFam, twenty landed below that line: the write-up is here.

07 Second demonstration

Vernier hyperacuity, measured in your browser

A separate instrument, from earlier work: the perceptual effect Qeben is built on, including an honest account of what a screen can and cannot measure.

Vernier acuity · adaptive staircase 2-down / 1-up · 14 trials · ~71% correct

This is a demonstration, not a vision test. It is not a medical device, it produces no diagnosis, and nothing here should inform a decision about your eyes. If your vision has changed, see an optometrist or ophthalmologist.

Step 1: Calibrate the display

Hold any bank card flat against the screen and drag the slider until the rectangle matches it exactly. Payment cards are standardised worldwide at 85.60 × 53.98 mm, which makes them a reliable ruler for finding your display's true pixel pitch.

Step 2: Viewing distance

How far are your eyes from the screen? Measure it if you can; a guess propagates straight into the arithmetic below.

cm

Step 3: The task

Two vertical segments appear with a gap between them. Decide whether the upper segment sits to the left or the right of the lower one. Fourteen trials, about a minute. The offset shrinks when you are right and grows when you are wrong, so it converges on your threshold rather than testing you at a fixed difficulty. Buttons or the ← → keys.

Angular size, logarithmic, arcseconds

60″Standard visual acuity (1′)
≈ 90″One screen pixel at 60 cm
2–5″Clinical vernier threshold
1″ 10″ 100″

Why a browser cannot measure hyperacuity, and why that is the argument, not the excuse.

Vernier acuity is the ability to tell that two line segments are not collinear. It is called a hyperacuity because the threshold is finer than the spacing of the photoreceptor mosaic itself: the retina does not resolve the offset, the cortex infers it from the luminance distribution across several receptors. Reported thresholds sit around 2–5 arcseconds, roughly ten times finer than standard visual acuity, which is about one arcminute, or 60 arcseconds.

That sensitivity is precisely why it is clinically useful. Metamorphopsia, the perceived warping of straight lines caused by fluid lifting the macula, shows up in a hyperacuity task long before it shows up on a letter chart. Preferential hyperacuity perimetry exploits exactly this, and has been reported discriminating wet AMD with an area under the ROC curve of 0.87.

Now put that on a screen. At a 60 cm viewing distance, a single pixel on a typical display subtends roughly 90 arcseconds. The quantity being measured is 2–5. One pixel is about twenty times coarser than the entire clinical range.

The demonstration above still reaches sub-pixel thresholds, and the reason is worth stating: an offset of a fraction of a pixel is not discarded, it is antialiased. The renderer distributes the line's energy across neighbouring pixels, shifting the luminance centroid by less than the pixel pitch, and the visual system integrates that shift. So the task is measurable, but below one pixel you are partly characterising the rendering pipeline, the panel's subpixel geometry and its gamma, not only the observer.

This is the distinction between a demonstration and an instrument. A browser can show you the phenomenon and let you feel where your own threshold lies. It cannot control viewing distance, luminance, contrast calibration, refraction, or pixel geometry, and without those the number is not a measurement of the eye. That is not a disclaimer bolted onto the end of a product. It is the reason the product was labelled not for diagnostic use, and it is the kind of constraint that has to be settled before any code is written, not after.

Sources: The Clinical Use of Vernier Acuity: Resolution of the Visual Cortex Is More Than Meets the Eye · Accuracy of a Self-monitoring Test for Identification and Monitoring of Age-related Macular Degeneration

08 In brief

David Mashiah, in one paragraph

The short version: who he is, what he's built, and what stands up to checking.

David Mashiah, supply chain and distribution expert in Tel Aviv

David Mashiah (Hebrew: דוד משיח; also written David Mashiach) is an Israeli supply chain and distribution expert based in Tel Aviv who brings AI and data analysis into logistics, and a mechanical engineering undergraduate at Tel Aviv University. He worked hands-on in perishable-goods distribution: seven months in flower supply, buying direct from the wholesaler and building to a costed spec, one of the most time- and spoilage-constrained supply chains there is — including the part where the product got three times better and sales fell. Alongside that he builds complete systems end to end, front end through deployment; the most visible, 1 vs 100, runs live for Tel Aviv University and holds up to 1,000 concurrent players in a single server process at zero infrastructure cost.

What connects the work is a single question: how do you know when a model is lying to you? The field that asks it is machine learning reliability, and his research measures exactly that. MetaXFam studies when machine-learning surrogates fail on inputs they were never trained on (the same failure that turns a demand forecast from slightly wrong into inverted the day conditions shift), and his coding-theory work applies finite-field Reed–Solomon error correction to sequence data, the same family of codes that repairs a scuffed barcode or QR label in a warehouse instead of only detecting the damage. Operating experience on one side, the mathematics of failure on the other.

The evidence is public. Every research output carries a DOI: a journal contribution on algebraic encoding of amino-acid sequences, a software toolkit for GF(2^5) Reed–Solomon error correction, the MetaXFam dataset, and MetaXFam-D, the benchmark that supersedes it and was rebuilt in August to remove a duplication flaw it found in itself. Three systems carry the engineering. Qeben, a home-monitoring research prototype for age-related macular degeneration, was engineered to the standards expected of regulated software but was never clinically validated and never deployed. 1 vs 100, built for Tel Aviv University and live at cup.taueng.com, holds a thousand concurrent players in a single room at no infrastructure cost. Leviathan is a local-first desktop application with no server at all.

This site is about a living supply-chain and distribution specialist in Tel Aviv who works with AI and data systems. The English name "David Mashiah" and the Hebrew string "דוד משיח" also occur as a religious term in Judaism, as the title of an Israeli television series, and as the name of other people, living and deceased. None of those are the subject of this site and it makes no claim to them.

The working notes are where the reasoning is shown rather than summarised — each one measuring where something breaks instead of reporting where it works. Most recent: Averaging threw the answer away before the model ever saw it, There was a random number hiding inside my measurement, and The bouquets got better and sales fell. All notes →

NameDavid Mashiah · דוד משיח · David Mashiach
FieldSupply chain & distribution · AI and data analysis
Based inTel Aviv, Israel
EducationMechanical engineering undergraduate, Tel Aviv University
LanguagesHebrew · Arabic · English
ORCID0009-0004-4684-955X
Entity recordWikidata Q140694415
Live systemcup.taueng.com · 1 vs 100, Tel Aviv University