RESEARCH3 days ago
The best config in your bake-off didn't win. Selection did.
Best-of-K eval selection bias: pick the highest-scoring config from K candidates on one eval set and...
News Feed
60articles · curated from global AI & Data Engineering sources
RESEARCH3 days ago
The US government publishes 7 million flight records a year. Nobody can read them.
Every delayed flight in America is documented in a public federal database. It arrives as a zip full of CSVs, so almost nobody opens it. Here is what happened when I did.
RESEARCH4 days ago
Zero failures isn't zero risk: the rule of three for evals
The rule of three for evals says zero failures in N runs is a count, not a rate. With 0 failures in N...
RESEARCH4 days ago
Your eval dashboard has 30 metrics. When one "moves," that is usually arithmetic, not a regression.
Here is the ritual. You ship a prompt change, rerun the eval suite, and open the dashboard. Thirty...
RESEARCH5 days ago
openFDA Adverse Event API: Paging Past 25,000 and Flattening
Measured guide to the openFDA adverse event API: beat the 25,000 skip cap with search_after, fix silent overcounts, decode code fields, flatten FAERS and MAUDE.
RESEARCH5 days ago
Your A/B eval is paired. Your stat test probably isn't.
Two prompts on one eval set are paired data. The two-proportion SE said 'collect more'; McNemar said it was already decided.
RESEARCH5 days ago
My trading system has a 73.5% win rate and loses money. Here is the diagnostic that found it.
I spent a few days auditing a live FX system I built, over 16 years of history. The result was negative in a way I did not expect, and the diagnostic
RESEARCH6 days ago
The Facebook URL Zoo: every page-link format and how to canonicalize them
If you scrape or analyze Facebook data, the same page will arrive under half a dozen different URLs —...
RESEARCH10 days ago
I turned a public-domain government CSV into a searchable index of every port and airport on Earth
For a side project I needed something that turned out to be weirdly hard to find: a free, clean list...
RESEARCH10 days ago
Fifty Poisoned Samples Is All It Takes
I spent this morning reading Anthropic's new research on how a tiny number of poisoned samples can...
RESEARCH11 days ago
Yes, another tree classifier. Here's what building one from scratch taught me.
I spent way too long building a tabular classifier from scratch. It doesn't beat XGBoost. I'm...
RESEARCH11 days ago
Small Models That Think Harder Beat Big Models That Sound Confident
People keep saying reasoning needs scale. That you can't get reliable chain-of-thought out of...
RESEARCH11 days ago
We graded 13 famous trading strategies on 6 years of data. All 13 failed.
Every trading book, YouTube channel, and Discord server teaches the same classics:...
RESEARCH13 days ago
Your embedding axes can move while cosine neighbours stay put
An embedding tour shows how an orthogonal basis switch changes the view while every top-five cosine neighbour stays fixed.
RESEARCH13 days ago
Our football model went 63-for-76 at the World Cup. Here are the 13 it got wrong.
Most football prediction sites publish a hit rate. Almost none publish the list of matches they got...
RESEARCH13 days ago
My Experiment Showed Zero Effect. A Statistician Told Me My Measurement Was Broken.
Last week, I ran an experiment that failed. The hypothesis was simple: syllogistic prompts ("Major...
RESEARCH14 days ago
What happens when you ask 8 AI models the same buying question every month
A while back I got annoyed at a specific genre of blog post: "we asked ChatGPT what the best CRM is...
RESEARCH16 days ago
I tracked every trending AI repo's stars daily for 3 weeks. The growth is not where I expected
I run a small AI trends site, and three weeks ago I started doing something simple: every day,...
RESEARCH16 days ago
A Fair Coin Isn't Enough: When a Perfectly Randomized Experiment Is Impossible to Analyze
Our A/B assignment used a cryptographically fair coin and was still unanalyzable — nothing joined the random draw to the task it dispatched, so no code could ever compute the intention-to-treat comparison. The join, the re-draw guard, and the power calc are the experiment; the coin is the easy part.
RESEARCH16 days ago
We Tried to Find a "Future Signal" in Bearing Failure Data. It Was Just Structured Degradation.
TL;DR: Pre-failure noise in degrading mechanical bearings is famously predictive of time-to-failure....
RESEARCH16 days ago
Tracking New Patent Filings Without an API Key (Google Patents Has a Hidden XHR Endpoint)
I wanted a simple thing: a weekly feed of new patent filings matching a keyword, so I could watch...
RESEARCH17 days ago
A CSV sample that looks like the whole dataset is worse than no CSV at all
I've been building data2prompt for a few months now. It takes a data-heavy project (CSVs, SQL dumps,...
RESEARCH17 days ago
One average eval score was hiding two different failure modes
A mean faithfulness of 0.75 sounds like a model that is usually right and occasionally slips. Mine...
RESEARCH17 days ago
One Master Dataset, Two Consumers: Building a Governed Pipeline for BI + ML
The problem with two datasets A recent project of mine called for two things that are easy...
RESEARCH19 days ago
How a JavaScript Pivot Table Can Handle 10 Million Rows in the Browser
A technical look at browser-local pivot table performance, columnar ingestion, Web Workers, TypedArrays, and the caveats behind a 10M-row benchmark.
RESEARCH20 days ago
Conformal intervals under a log transform: the blow-up isn't a back-transform bug
Someone log-transforms their target, fits a model, wraps it in conformal prediction,...
RESEARCH22 days ago
Building a world-history map: PostGIS time+bbox search, a full-screen MapLibre UI, and Wikidata ETL lessons
Annalium is a free, no-login map + multi-track timeline of world history. You pan a full-bleed world...
RESEARCH24 days ago
Episia: Building an Open-Source Epidemiology Library in Python, Validated Against OpenEpi
Why Episia I am an MD candidate and a self-taught Python developer. Over the past year,...
RESEARCH24 days ago
Expected goals (xG) from scratch, and why the scoreboard lies
Argentina lost their 2022 World Cup opener to Saudi Arabia, and everyone said Saudi were the better side. They weren't, not even close, and one number proves it. Here is expected goals built from scratch in about 150 lines of Python, honest about where it is brilliant and where it lies.
RESEARCH25 days ago
How to Build an H-1B Salary Database by Employer (the Real Data Source + Python)
A practical guide to querying H-1B salary data by employer and role programmatically, using the authoritative DOL OFLC LCA disclosure files with runnable Python.
RESEARCH26 days ago
Stop Writing Python Loops: Vectorization and Broadcasting in NumPy
In production systems, the difference between a Python loop and a vectorized NumPy operation isn't...
RESEARCH27 days ago
Your p-value answered a question you didn't ask.
You ran the A/B test. It came back p = 0.08. "Not significant." So you killed the feature and moved...
RESEARCH27 days ago
samkhya v1.0: Plug Claude, GPT-4o-mini, or Local Ollama Into Your SQL Query Optimizer
samkhya v1.0 ships an LLM-pluggable corrector backend for embedded analytical engines— DataFusion, DuckDB, Polars, Postgres, Iceberg, gpudb. Plug Clau
RESEARCH28 days ago
I built a dataset of 50,000 debugging sessions — and what I found surprised me
Most bug datasets only tell you: "Bug was fixed in 23 minutes." They don't tell you what happened...
RESEARCHabout 1 month ago
Generating Synthetic Enterprise Datasets for AI Systems
Part 2 of the Building Enterprise AI Automation Systems Series ...
RESEARCHabout 1 month ago
Using Zstd Frames to Egress Partial Parquet Files
Jump Tables, TLV Footers, and the Real Cost of Reading What You Don't Need You're paying...
RESEARCHabout 1 month ago
How we predict the FIFA World Cup 2026 with a Dixon-Coles bivariate Poisson model
We're building Onside Arena — an open AI football analytics platform for the FIFA World Cup 2026 and...
RESEARCHabout 1 month ago
Stop Funding History: Building a Demand Prediction App to Optimize Trade Spend
Let’s be honest: in the FMCG (Fast-Moving Consumer Goods) space, trade marketing budgets are almost...
RESEARCHabout 1 month ago
When SuSiE Says '95% Confident', Is It?
Benchmarking the Honesty of Fine-Mapping Credible Sets Fine-mapping has a promise built into its...
RESEARCHabout 1 month ago
Your Variant Caller Tells You How Confident It Is. Have You Ever Checked If It's Telling the Truth?
Every variant caller you have ever used attaches a number to each call: QUAL. It is a confidence...
RESEARCHabout 1 month ago
The App Store's silent giants: AI assistants reply to almost none of their reviewers
An App Store rating looks like a verdict. It behaves more like a monument, built over years and slow...
RESEARCHabout 1 month ago
How I Replicated Uber's Core Marketplace in Python: A Technical Deep Dive
I Built UberSim v2.0: A Production-Grade Urban Mobility Intelligence Platform 🚗🧠 Every...
RESEARCHabout 1 month ago
I built a 200-page programmatic SEO site. Bing out-ranked Google 2.6X .
Ten weeks ago I launched a niche reference site on a fresh .in domain. No backlinks, no budget, no...
RESEARCHabout 1 month ago
Your bandit's exploration floor probably violates its own floor
Most multi-armed bandit / A-B allocation systems add a minimum exploration weight: every arm should...
RESEARCHabout 1 month ago
When your optimizer silently returns the wrong answer (and how to catch it)
Numerical solvers have a failure mode that is worse than crashing: every so often they return status:...
RESEARCHabout 1 month ago
Deploying Label Studio Open-Source Data Labeling Platform on Ubuntu 24.04
Label Studio is an open-source data labeling platform that supports annotation across text, images,...
RESEARCHabout 1 month ago
Stop Shipping ML Models With Bare Floats: A Deep Dive Into Statistically Rigorous Model Evaluation
Stop Shipping ML Models With Bare Floats Every week, somewhere, a team makes a deployment...
RESEARCHabout 1 month ago
The Disk-Level Architecture of OLTP vs. OLAP
Every backend engineer has seen this happen, you build an application on a relational database like...
RESEARCHabout 1 month ago
Building KNN from Scratch (Because import sklearn Feels Like Cheating)
Building K-Nearest Neighbors (KNN) From Scratch Let's be real: in a production machine...
RESEARCHabout 1 month ago
The SEC EDGAR API: A Practical Guide to Free Filing Data in Python
A hands-on guide to the free SEC EDGAR API: the required User-Agent header, ticker-to-CIK mapping, XBRL financials, and full-text search in Python.
RESEARCHabout 1 month ago
Why your synthetic fintech data fails code review (and how mixture models fix it)
Every fintech developer has done this: you need test data, you reach for Faker, you generate ten...
RESEARCHabout 1 month ago
We put confidence intervals on our LLM-judge scores. The error bars ate three weeks of "trend"
We track weekly agreement between an LLM judge and human labels (Cohen's kappa) on a sample of...
RESEARCHabout 1 month ago
I Built an LSM-Tree in Pure Python and Finally Understood How RocksDB and Cassandra Handle Millions of Writes
I have built a write-ahead log and a bloom filter from scratch in the last week. Today I wanted to...
RESEARCHabout 2 months ago
The Anatomy of Catastrophic Forgetting
We train a model on handwritten digit classification. 99% accuracy. Then we train the same model on a...
RESEARCHabout 2 months ago
When Feature Importance Lies: Target Encoding at the Noise Floor
We built a feature that dominated our model's importance rankings. Across four random seeds and three...
RESEARCHabout 2 months ago
I Shipped 12 BI Dashboards With 5 Different Tools. Here Is the Honest Comparison.
Streamlit, Plotly Dash, Apache Superset, Evidence.dev, and Grafana -- not from a weekend evaluation, but from 12 production pipelines. Here is when each one earns its place and when it gets in the way.
RESEARCHabout 2 months ago
How I Used Python to Analyse 40,000 Human Gut Cells and Uncover What Makes Crohn's Disease Different from Colitis
A step-by-step walkthrough of my multi-sample single-cell RNA sequencing project — written for anyone...
RESEARCHabout 2 months ago
How I Mapped Brain Cell Changes in Alzheimer's Disease Using Single-Cell RNA Sequencing
Alzheimer's disease affects over 55 million people worldwide, yet the precise molecular changes...
RESEARCHabout 2 months ago
I Built a Fraud Detection System That Catches 99.76% of Fraud — Here's Everything I Learned
There is a number that haunts every fraud detection engineer: 0.13%. That is the fraud rate in the...
RESEARCHabout 2 months ago
Machine Learning for Data Engineers: The Patterns I Actually Used Across 7 Projects
Not a data scientist, but I've shipped ML models in production across 7 projects. XGBoost regression, 109 Prophet forecasts, SHAP explainability, FinBERT sentiment, and pgvector embeddings -- here's what the stack looks like from a DE's perspective.