Selected work

Systems that had to be correct.

Every engineer here practiced structural, civil, or manufacturing engineering before they wrote software. These are the systems that came out of that order, where knowing the discipline changed the architecture and not just the vocabulary.

We do not name clients by default. Ask, and we will take you through any engagement here in detail under NDA, including who, and the engineers who did it.

Defense · Agentic AIProgram leadActive

Design review is a gate at the end. We moved it into the drawing.

Military construction review happens after the fact. Architects work for months, submit, and wait. Only then do they learn which provisions of the Unified Facilities Criteria and the International Building Code they missed. Every issue found at that point is expensive. Every issue found on site is worse.

We serve as principal investigator and lead architect on the federal program behind it. The system encodes national and defense construction codes as a queryable knowledge graph, reads the drawings as well as the model, and checks a live building model against them while the architect is still drawing. Reviewers can run it from inside the tools they already work in.

The call we made

A compliance answer that sounds right is worth nothing. The bar is whether a licensed professional will sign against it. So the AI gathers the evidence and deterministic code reaches the verdict, a second pass argues against our own findings before anything reaches a person, and every conclusion opens to the reasoning that produced it.

The rule we hold it to

The AI gathers the evidence, deterministic code reaches the verdict, and every conclusion shows its work.

Publicly awarded Phase II SBIR ↗Every release is checked against a bank of 80 known-correct code answers before it ships.
Read the full engagement →

Proving what got built matches what was designed.

DefensePrincipal investigator · Delivered

A federal program to automate the design, construction management, and accreditation of secure facilities, where an accreditation decision rests on the finished room actually matching the drawing. We led the team behind a graph-based analysis engine for building models, an iOS augmented-reality app for checking a design at life size on site, and a point-cloud scanner that reconstructs a finished interior in three dimensions.

Field verification only counts if its tolerance is tighter than the thing it is checking.

Publicly awarded Phase II SBIR ↗A quarter inch over thirty feet.

A wall is not a picture of a wall.

Construction roboticsFounding engineer, then contracted back · Ongoing

Turning an architectural model into the path a robot follows sounds like a graphics problem. It is not. Every offset, every arc approximation, every place two walls meet is a structural decision wearing geometry’s clothes, and the robot pours the consequence in concrete. We built that engine, then were brought back years later to mature it as the buildings got harder: curved forms, multi-story reinforcement, code-governed splice geometry.

The call we made

Rebar splice selection was being decided by bare floating-point comparison. Two lengths that agree within tolerance are the same length to a structural engineer and a different answer to a computer. We replaced those comparisons across the reinforcement pipeline. A generalist reads that as a rounding bug. A structural engineer reads it as a splice that might not develop.

200+ homes printed across active sites, including a U.S. Army barracks program.

No cloud. No vendor. No signal.

Defense · roboticsEngagement lead · Active

Construction robots deployed where there is no connection back to anyone. A job has to load while connected and then run to completion long after that link is gone. We designed and built the execution path that makes it possible.

The call we made

We set one rule before writing any of it. A job package that cannot be assembled completely fails outright rather than arriving quietly incomplete. A crew in a field with no signal cannot recover from a package that only looks whole.

A crew can trust what it loaded before it starts a print.
Deep tech · computational designEmbedded architectActive

Most CAD files record a finished shape. This one records the intent that produced it.

A company is rebuilding CAD from first principles. Its format carries the parameters, materials, objectives, and physical constraints behind a design, with geometry written as differentiable fields. That closes a loop the industry has never closed: an optimizer stops searching a catalogue of variants and starts moving the design itself. Drawing something and solving for it become the same operation.

We came in to help take that from research code to a platform outside engineers could build on, and stayed embedded in the team. The decisions we sit on now are the ones that outlast us: the schema and versioning contract for the core format, the architecture of the platform around it and where each part of it runs, and where the representation itself should go next, from kinematics to coupled physics to tolerance-aware design.

Why it matters

A file format is a set of claims about what the physical world is allowed to be. Decide it narrowly and every product built on top inherits the ceiling. That decision needs someone who has designed real parts, not only written parsers.

The shift, in one line

A file that stores a shape becomes a file that stores the argument for the shape.

When a company is inventing the file format, someone has to decide what it can and cannot say. We are in that room.
Read the full engagement →

They could build the system. They could not tell if it was right.

Municipal permitting · AIAdvisory · Active

A startup checks construction plans against a city’s regulations and returns what passes, what fails, and what needs clarification, aiming to take weeks out of a permitting cycle. Their engineers were good. But no amount of prompting tells you how a plan reviewer actually reads a sheet, or what separates a finding that holds up from one that gets waved away.

That gap is the entire engagement. We supplied the judgment: how a review decomposes into steps an agent can execute, what makes a finding defensible, and where their output diverged from what a practising reviewer would accept.

This is the whole thesis in one engagement. The code was never the missing piece.

We tried to break our own findings before the client saw them.

Venture-backed productTechnical discovery · Delivered

A venture-backed team was deciding whether to put more money into a product their own engineers could no longer fully account for. They needed to know what they actually had, across four repositories: architecture, mobile readiness, backend and data design, security, regulatory posture, and every screen in the app.

The call we made

Most diligence stops at a list of concerns. Ours ran twice. A second, independent pass went back to the source code with one job: falsify the first pass. What survived went in the report. What did not was cut before the client ever read it.

A launch decision in weeks, with a phased plan, priced options, and the open questions named rather than buried.
Civil infrastructureEmbedded product teamIn build

The bottleneck was never the digging.

A utility contractor digs the holes for utility pole replacements. Crews and machines were never the constraint. Before anyone breaks ground, a job has to clear underground-utility locates, environmental and cultural-resource permits, traffic-control plans, and inspections, and every one of those has an expiry date. Clearance is what sets the pace of the business.

Run on paper and spreadsheets, that is not a scheduling inconvenience. A lapsed locate that nobody noticed is a crew digging without authorisation, and the exposure that follows lands on the contractor, not the spreadsheet.

We are building the system that replaces it: job and audit tracking, scheduling that understands clearance, a live map of every excavation point, and field progress logged from the truck rather than typed up that evening.

The call we made

The scheduling gate lives in the database, not the application, so no screen and no future feature can route around it. And guarding a gate is pointless if its inputs stay editable: a lapsed ticket whose dates can be bumped is a one-line bypass of the whole model. So the dates a clearance rests on freeze the moment a job passes, and a correction takes an administrator who has to record a reason.

The property we designed for

A release gate is only as strong as the mutability of its inputs.

There is no path through the software that can schedule an uncleared dig. Not a warning, not an override anyone can click. The database refuses.
Read the full engagement →

Grading that used to depend on which inspector was looking.

EnergyApplied machine learning · Delivered

External corrosion on offshore oil and gas structures gets graded from inspection photographs, by eye, by whoever is on the job that day. We trained a computer vision model to grade it instead, at a global engineering firm.

Knowing what a grade means structurally is the difference between labelling images and assessing damage.

If one of these looks like your problem, the conversation starts with your system, not our deck.

We will read your code, talk to your team, and tell you what we actually think before anyone signs anything.