Vol. XII · No. 05 · May 2026
Jake Cuth.

I build data systems that ship.

The kind that pay for themselves. Models that keep the accounts worth keeping, aim spend at the customers who convert, and stop losing tests before they cost a quarter. Forty-five live labs below, filterable. Click anything.

Hancock Whitney Nissan ReminderMedia Comcast NBCUniversal Warner Bros. Georgia Tech Temple University

Forty-five labs.
Hover any tile.

Color tells you what kind of work it is. Hover for the details. Click to open. Filter by category, or search by topic.

45 labs
No labs match. Try a shorter query or a different category.
M01

Churn Model

Finds the customers about to leave. Saves revenue already paid to acquire.

3 models · ROC overlay · fairness auditOpen →
M02

Customer Segmentation

Sorts customers by value. Aims spend at the 21% who generate 64% of revenue.

Drag the cost · live recomputeOpen →
M03

Recommendation Engine

Guesses what they'll watch next. Every basis point of watch-time is a subscription renewed.

3 reco models · same likesOpen →
M04

A/B Test Simulator

Shows when a winning version is real and when it's noise. Stops losing tests early.

Bayesian + frequentist · priorsOpen →
M05

Funnel Simulation

Finds the one customer segment quietly killing conversion without breaking the rest.

4 levers · monte-carlo · segment-awareOpen →
M06

Sketch Recognition

Draw anything. A small neural network running on your laptop guesses what it is, in real time.

30 classes · ~930KB ONNX · in-browserOpen →
M07

Model Picker

Ask in plain language: "cheapest model good at code." Get a side-by-side comparison.

300+ models · OpenRouter · weeklyOpen →
M08

Influencer Mix Modeling

A working Bayesian Marketing Mix Model for a creator-driven D2C brand. Eight channels, 104 weeks, full posterior intervals.

8 channels · 104 weeks · Meridian on GCPOpen →
M09 · NEW

Token Lab

A live Byte Pair Encoding visualizer. Type text, watch merges happen step by step, see exactly how GPT turns words into numbers.

Pure JS · BPE algorithm · 50K vocabOpen →
M10 · NEW

Bandit Lab

A live multi-armed bandit simulator. Four strategies, hidden reward probabilities, and real-time regret curves that prove why exploration matters.

Pure JS · Monte Carlo · Beta posteriorsOpen →
M11 · NEW

Diffusion Lab

A toy 2D diffusion model in your browser. Draw a shape, watch it dissolve into noise, then walk it back step by step. The engine behind DALL-E.

Pure JS · Canvas 2D · DDPM toyOpen →
M12 · CASE

Viral Nation Case Study

Interview deliverable for Viral Nation. Hosted externally on Vercel. Opens in a new tab.

External · vn-case.vercel.appOpen ↗
S01

AI & Jobs

Type your job. See its AI exposure, robotics exposure, and BLS projection through 2030.

888 occupations · 3 measuresOpen →
S02

Time-Use Atlas

How Americans actually spend 24 hours, drawn as a clock. Pick a cohort, watch the day redraw.

13 cohorts · 13 categories · ATUSOpen →
S03

Algorithm Watchdog

A daily-refreshed read on what YouTube's algorithm is rewarding right now.

30 days · 50 videos/day · 11:00 UTCOpen →
S04

AGI Horizon

Two hundred fifty AGI predictions plotted on year said vs year targeted. Twenty years out, since 1965.

250 predictions · 1950 → 2026Open →
S05

The Productivity Mystery

U.S. labor productivity broke in 1973 and again in 2004. Seven datasets, seven explanations, one survives.

8 charts · 1947 → 2025 · G7 panelOpen →
S06

How LLMs Learn

A trillion words, a handful of equations. Eight charts on what actually happens inside an LLM.

8 charts · Vaswani · Kaplan · SchaefferOpen →
S07

Escape Velocity

Are we approaching the singularity? Seven charts. The inputs are bending up. Is intelligence following?

7 charts · 11 datasets · 1950 → 2030Open →
S08

Semiconductor Cartography

Thirty facilities, seven supply-chain layers, the chokepoints that could stall the AI chip industry.

30 nodes · 7 layers · 36 connectionsOpen →
S09

Tech Predictions That Never Happened

Eighty-two confident, dated technology predictions. Each named a year. The year came. Most did not.

82 predictions · 1940 → 2026Open →
S10

The Productivity Paradox

Ninety-one predictions about when IT and AI would show up in productivity statistics, against the BLS series.

91 predictions · 1978 → 2025 · BLSOpen →
S11

The Skills Surge

Six months of Claude Skills, mapped. Power-law star distribution, surge timeline, architecture quadrant, and the verbs every skill author opens with.

4 charts · 117k★ · 17 official · 22 vendor packsOpen →
S12

Inside Agent Memory

The 2026 map of agent memory. MEMORY.md vs vector DBs vs Mem0 vs Zep vs the hybrids in between, with benchmarks, costs, and a decision matrix.

4 charts · 12 approaches · LongMemEval 49 → 95Open →
S13

Cosmic Engines

An interactive 3D field guide to the eight extreme objects that light the universe. Quasars, black holes, pulsars, and the violent machinery between.

8 engines · scroll-driven WebGL · NASA-sourcedOpen →
S14 · NEW

Reservoir Teacup Map

Twenty Western reservoirs drawn as USBR teacup diagrams. Current storage versus historical average, plus a Mead/Powell elevation cliff and a snow-to-storage scatter.

20 reservoirs · 5 basins · SVG small multiplesOpen →
S15

Neurotransmitter Atlas

A chapter-style 3D neuron, synapse close-up, and brain map for six major neurotransmitter systems.

6 transmitters · 6-step signal path · rotatable 3D specimenOpen →
S16

Better?

A scrubbable ledger of one hundred indicators of human progress and regress, from primary sources.

100 indicators · 5 themes · 1820–2025Open →
S17

Bubble Watch

Six charts on whether the U.S. market is in an AI bubble. CAPE, concentration, hyperscaler capex vs frontier-lab revenue, margin debt, capability scaling, and a five-peak comparison.

6 charts · 11 datasets · 1881 to 2026Open →
S18

The Price of Intelligence

Eleven commodities, seven centuries. A stress test of the claim that AI is the first commodity with infinite demand. 1,000x in 3 years for intelligence vs 12,000x over 700 years for light.

5 charts · 11 commodities · 1300 to 2026Open →
S19 · NEW

The Rework Tax

Eight datasets stress-test the viral claim that 82% of AI tokens go to rework. The direction is right. The number is marketing. The baseline is from 1978.

8 charts · 8 sources · 2020 to 2026Open →
P01

Inside DeepSeek-V4

Five things most people assume modern LLMs do, and what DeepSeek-V4 does instead.

5 paired comparisons · 27% FLOPsOpen →
P02

Inside Subquadratic

A real Miami AI startup, $29M seed, $19.6M GPU contract, benchmark numbers no one has reproduced.

9 claims scored · 12M contextOpen →
P03

Attention Is All You Need

An interactive infographic of the 2017 Transformer paper. 6×6 attention heatmap, 8 heads, complexity slider.

2 interactives · 28.4 BLEU EN-DEOpen →
P04

Inside LeWorldModel

The 15M JEPA world model that replaced seven loss terms with one Gaussian-matching regularizer.

2 interactives · 192-d CLS · 48× planningOpen →
ATLAS

The Model Atlas

An expert system over 13 ML models. Five questions, one match, with reasons.

13 models · 5 questionsOpen the Atlas →
A01

Linear Regression

Fitting a line. Drag points, watch coefficients update in real time.

Live demo · from the AtlasOpen →
A02

Logistic Regression

The simplest classifier worth deploying. Plant points, watch gradient descent slide a probability boundary into place.

3 weights · live SGD · 0 depsOpen →
A03

Ridge & Lasso

Two regularizers, side by side. Watch coefficients shrink, and Lasso zero them out.

Two regularizers · live shrinkageOpen →
A04

Decision Tree

Recursive splits made visible. Depth slider shows over- and underfit.

Depth slider · partition + treeOpen →
A05

Random Forest

A committee of decision trees. Many imperfect trees beat one perfect one.

Many trees · one good answerOpen →
A06

Gradient Boosting

Each tree fixes the last tree's mistakes. The technique behind XGBoost.

Sequential trees · residual fitOpen →
A07

Support Vector Machine

The widest margin wins. Soft-margin linear SVM, live SGD on hinge loss.

Hinge loss · live SGDOpen →
A08

Multi-Layer Perceptron

Layers of nonlinearity. Live backprop, one batch at a time.

Live backprop · one batch at a timeOpen →
A09

Naive Bayes

Bayes' theorem applied bluntly. Wrong assumption, surprisingly useful answer.

Conditional independence · liveOpen →
A10

K-Nearest Neighbors

No training. Just memory. Predict by voting among k neighbors.

k slider · 3 datasetsOpen →
A11

K-Means

Lloyd's algorithm, drawn live. Pick k, watch points snap to centroids.

Live iteration · pick kOpen →
A12

DBSCAN

Density-based clustering, no k needed. Eps and minPts control everything.

Eps + minPts · live clusterOpen →
A13

Isolation Forest

Outliers are easy to isolate. Random splits, short paths to anomalies.

Random splits · live scoreOpen →

Dirty data,
clean production.

End-to-end builds, not slideware. Each one shipped with the plumbing, the UI, and the deploy. Usually that's three teams and two quarters of work. Below is what happens when one person owns it.

Personal · Agentic

Claude Second Brain

A note graph where Claude backfills links, summaries and open questions. Runs locally against a vault.

ReceiptsRuns locally · vault-agnostic · backfills links + summaries
Claude Code · Markdown
Hardware · Dashboard

Open Claw Mission Control

A control surface for a DIY claw machine. Live camera, queueing, telemetry and a leaderboard.

ReceiptsLive camera + queue · web telemetry · physical leaderboard
Linux · VPS · Python · JavaScript
Delegator screenshot
Delegator
iOS · Tasks

Assign work to your future self. Re-asks you at the right times instead of at the wrong ones.

Stillhue screenshot
Stillhue
iOS · Generative AI

Extracts a 5-color palette from any photo with a deliberately slow, gestural picker.

Second Order screenshot
Second Order
iOS · Decisions

A journal that flags second-order consequences before you commit to them. Prompts pull from your own past entries.

Sep 2023 — May 2026 (expected)

Georgia Institute of Technology

M.S. in Analytics
Sep 2014 — May 2019

Temple University

B.B.A. in Economics · Fox School of Business

Plus 25+ certifications: machine learning (U. Washington), data science (CU Boulder, Duke), people management (Google). full list on linkedin ↗

Let’s build
something.

Best first move: send the job and the numbers it needs to hit. I reply within two business days with relevant examples and code.

215 · 834 · 2151  ·  Philadelphia, PA  ·  39.9526° N, 75.1652° W