01
refi-leads
Scores multifamily refinance timing and finds the owner behind the entity.
Commercial real estate finance
Live since June 2026, in beta
A multifamily loan's maturity date is public years before it comes due, so knowing when to call is the easy part. Reaching the owner is harder, because title usually sits with an LLC and an LLC has no phone number. The ranking has to be defensible as well, since a broker will ask why one borrower came up above another, so it is built from just three signals: maturity, LTV headroom, and portfolio size. Finding the person behind the entity runs as a background job against a skip-trace service, with public filings as a fallback when a public company owns the property.
~10,800Lines of TypeScript
190Tests
3Scoring signals
Read the refi-leads case study →
02
Wellbore premium allocation
Turns an unauditable premium spreadsheet into a platform with a contract.
Energy insurance
Live since August 2026, in beta
Insurance premium for a schedule of wells has to be split across four coverages: Control of Well, General Liability, Umbrella, and Excess. The spreadsheet this replaced produced numbers nobody could fully audit, and the reason turned out to be a single column in the source data. That column decided whether a well qualified for General Liability at all, but it never appeared in the output, so you could see the result without ever seeing what drove it. The platform puts that reasoning back on screen, and it holds itself to a hard rule: re-running the 83-well reference file has to reproduce every figure exactly. If a change turns those tests red, the change is wrong until someone proves otherwise.
83Golden reference wells
24Test modules
4Coverage types
Read the Wellbore case study →
03
Sous
An agent that plans, asks, and adapts, entirely over email.
Autonomous agents
Running daily
Every Sunday morning Sous emails five dinner options, and the reply comes back the way people actually write: drop 2 and 4, more chicken. Turning that into something a program can act on is the real work, because it has to become two removals by position plus a preference applied to whatever fills the gaps. Once the plan settles it builds a grocery list grouped by aisle, keeping a running total that has to land inside a budget range, where coming in under the floor counts as a miss too. Weeknights it sends that evening's recipe. The whole interface is email, and model calls go through a CLI, so there is no API key to hold.
~2,600Lines of Python
9Test modules
0API keys
Read the Sous case study →
04
Recall Guardian
Five federal recall feeds, one schema, one severity scale.
Consumer safety
Ingestion live, app in progress
Five federal feeds publish product recalls, and no two of them describe a recall the same way. The field names differ, and so does the language for danger: two assign a class, one issues a park-it advisory, and two publish no severity at all. A parent trying to find out whether a car seat was recalled would have to check all five, which nobody does. The ingestion engine polls them on a staggered schedule, normalizes everything into one schema, assigns a severity level, and drops duplicates on a source-and-ID constraint. That half runs today; the consumer app on top of it is still being built.
728Recalls ingested
5Federal sources
4Severity levels
Read the Recall Guardian case study →
How I work
The specification gets written before any code does. It names the constraints, the decisions
already settled, and the things the system will deliberately not do. That last list is usually the
most useful one. On all four of these, most of the thinking happened before the first file existed.
That document is mine to write, along with the tests and the review gates. AI agents write the
code against all three. The difficulty moves rather than disappears. Code arrives faster than it used to, and so do
wrong answers, so the work is in knowing exactly what should exist and in catching output that looks
correct and is not.
Golden tests are treated as the contract. Where a system already has a known-correct output,
that output becomes the test, and the burden of proof inverts: the reference is presumed correct, and
a change that turns it red has to argue its way out. A coverage percentage tells you none of that;
a number that has to match is a specification you can run.
Completed work gets an adversarial pass: a separate review whose only job is to find what the
build missed. On refi-leads it found a task endpoint that accepted a client-supplied assignee
without verifying workspace membership. One tenant could have written into another tenant's data. It
was caught before launch and fixed by scoping every write to the workspace, then re-verified. I keep
that pass because reviewing your own finished work honestly is close to impossible.
Every session closes with a handoff document. It records where things stand, what is still open,
and what was left alone on purpose. A cold session reads it and keeps going. It is also the only
reason I can put a project down and pick it back up without re-deriving the reasoning.