the gallery
Every card is a thing that silently breaks in real data systems. Every demo runs in your browser.
One method at two scales: a deep case study (an agentic MLOps platform, built and measured end to end) plus small free-standing prototypes, each reverse-engineering one company’s publicly stated data-systems bottleneck and solving a slice of it, honestly measured. The same standard of evidence in both, so the reader sees one discipline applied at two densities, not a pile of demos.
Data can go bad without a single error. The table still looks valid; the numbers are wrong.
Flip a corruption on, watch the guardrail catch what the schema contract passes.
6/6 vs 1/6 silent corruptions caught on lerobot/pusht · zero false positives · 1.8 ms/check
Physical Intelligence · data core
A model can train on answers from the future. Offline accuracy rewards the leak.
Toggle the leak on, watch the score jump to a number too good to trust.
0.999 leaked vs 0.771 correct on the same 1200-row training set · guardrail ~0.1 ms
feature stores · data core
A cache can pass every functional test and never hit once. Every request pays full price.
Break the cache, watch cost climb while correctness stays green.
95% hit ratio working vs 0% broken over 240 requests on 12 hot prompts · 20× redundant compute
prompt caching · throughput
A query can return the right rows while reading everything. Row checks stay green either way.
Switch the scan mode, watch the byte counter disagree with the row counter.
44.7× the bytes for the same correct answer, 200k rows, 2 of 12 columns needed
columnar formats · storage
A feed can change meaning without dropping a row. Revenue reads $157k when the truth is $299k.
Drift the units mid-stream, watch the contract object while schema and null checks pass.
contract 6/6 drifts caught vs schema/null checks 1/6 · zero errors, zero nulls in the bad data · ~0.1 ms/check
ELT · schema-drift
An order gets nanoseconds to be checked before the market moves. Too slow loses the trade; too lax loses the firm.
Stream a fat-finger session through the real Rust crate, compiled to wasm, running in your browser.
37ns full 7-check evaluation · 23M evals/sec native (Criterion, repo hardware) · 3.3KB wasm build
electronic trading · execution core
Denoising an image one pixel at a time is 786,432 numbers a thousand times over. The word latent is the trick that made it fit on a laptop.
Drag the timestep, watch a field dissolve into noise and come back, then break the noise predictor by 5 percent.
48x fewer values to denoise at f = 8, computed live from the repo’s autoencoder config · schedule and shapes cited to file, no trained model on the pageThis unit shows the mechanism. It carries no measured result.
CompVis latent diffusion · mechanism unit, outside the bottleneck spine
A clinician talks for four minutes. The chart needs five headings, and every one of them has to be supported by something that was actually said.
Feed it a thin transcript and see whether the note says not documented or quietly invents a diagnosis.
A real model on Cloudflare’s GPUs, called live from this page · synthetic transcripts, nothing stored, no evaluation claimedOne model answering one transcript. No accuracy number, no hallucination rate, and not a medical device.
Applied AI · live inference on the free tier
A company that runs on mainframe files and CSV drops, taken to the point where an AI agent can make a business decision on it, safely.
Watch the engine take four mock businesses from legacy mess to a policy-gated agent decision, rendered from its own runs.
4 packs with planted defects and ground truth · write-audit-publish on Iceberg · hash-chained audit · $0, local, no keysThe films are the engine's own runs, and the demo verifies the run's hash-chained audit log in your browser. No pack executes end to end on this page yet.
Legacy modernization · one engine, four mock businesses
The deep one, still being built. One platform proving one claim: the classic MLOps loop still holds when the thing served is an LLM agent, once you add a token/cost axis and a reasoning-trace artifact.
Take the guided tour: three acts, seven stops, then explore all fourteen lifecycle stations.
the deep case study · ten units
coverage · one bottleneck class per row