The science of better decisions, in plain English
Field notes on operations research, optimization, and AI — grounded in real work and published research, written to be useful before it's clever.
Optimization over language: an optimizer that reads its rules in plain English
Classical optimization can only work on numbers, so for decades the objective and constraints that lived in prose — a preference, a policy, a note — got thrown away or hand-coded. A calibrated decision model changes that: it turns text into trustworthy coefficients and constraints a solver can use. I built a reviewer-assignment engine to test it, and it did something no similarity system can — it honored rules written in plain English, while matching experts better than a real embedding.
ReadMaking our own AI honest: a calibrated groundedness gate for RAG
A retrieval assistant answers only from your documents — until it quietly doesn't. We pointed a second, calibrated model at our own assistant to ask one question of every answer: is this actually supported by what we retrieved? On our own content it caught 93% of subtle hallucinations and every wrong-context answer, and refused none of the good ones. So we shipped it — fail-open, so it can never take the assistant down.
ReadConfident is not calibrated: moderation with a model that means its numbers
Any language model will tell you a comment is toxic. The question that decides whether you can automate anything is different: when it says 0.8, is it right 80% of the time? We ran a calibrated model over 1,500 real, human-labelled comments to find out — and to see how much of a moderation queue you can safely take off a person's desk.
ReadWhat gets fixed first: triaging a real 311 backlog by reading the text
A city's 311 backlog is always bigger than its crews, so every day is a triage. I gave a calibrated AI the citizens' own words, let it read each report for danger, and routed a fleet on real road times — then compared it against the rule most operations actually use: oldest-first. The result has a caveat worth stating and a punchline worth sharing.
ReadWhat the credit score can’t read: 1,500 real loans and one calibrated model
I gave a new kind of AI — a calibrated decision model — nothing but the words and numbers on 1,500 real loan applications, and hid the bureau’s own risk grade. The raw answer was wrong. Then it got interesting: once recalibrated, the borrower’s own words separated defaulters the credit grade could not — most sharply inside the “safest” band.
ReadThe math behind a smaller fleet: how optimization removes vehicles without cutting service
Most fleets are sized by experience, and experience over-provisions. Here is the operations-research reframe that removes vehicles while every trip still runs on time.
ReadOperations research or machine learning: which one actually makes the decision?
Everything gets called “AI” now. But prediction and decision are different jobs — and picking the wrong tool is why so many AI projects stall before they change anything.
ReadDecisions that survive contact with reality: a plain-English guide to robust optimization
The “optimal” plan often fails, because the numbers it was built on were never certain. Robust optimization is how you decide well when the inputs can move.
ReadDemand forecasting a retailer can actually plan on
A single-number forecast is a trap. Real forecasting gives you a range, is honest about what it doesn’t know, and only pays off when a decision hangs off it.
ReadAfter the buses: cutting legal driver duties and a fair roster
Finding the minimum fleet is the easy half of transit optimization. Turning those vehicle blocks into legal driver duties and a fair weekly roster is the hard, expensive, human half.
ReadGrounded AI: real answers from your documents, without the hallucinations
In the enterprise, a fluent wrong answer is worse than no answer. Grounded AI ties every response to your own material — with citations, and the discipline to say “I don’t know.”
ReadPredict, then optimize: why the most accurate forecast is not the best decision
Almost every operational AI is really two models in a trench coat — one that predicts, one that decides. Train them separately and you optimize the wrong thing. Here is why, and what to measure instead.
ReadEngineering for everyone: software the least tech-comfortable person can still use
Accessibility is usually a finishing pass. Building a free church finder for Brazil — where many users are on older Android phones, live inside WhatsApp, and are new to apps — meant treating it as architecture from the first commit.
ReadThe math of balance: what a breakthrough in discrepancy theory says about the decisions we optimize
In late 2025, two computer scientists cracked open a 40-year-old conjecture about spreading things out as evenly as possible. It won’t route your fleet tomorrow — but it is a clear window into the science underneath every decision we optimize.
ReadFive operations-research signals from 2026 — and where BIS already stands
Optimization is having a moment in the industry press: routing, crew scheduling, and resilience are on the front page. Here are five 2026 signals, and the plain reason each one plays to a firm built on optimization science.
ReadWhen the AI writes the code: engineering a data analyst that runs real Python, safely
A new class of tool answers questions about your data by writing and executing real code, not by guessing. The chat is the easy part. Here is the engineering underneath — sandboxing untrusted code, grounding every number, and letting an agent drive the whole thing.
Read