Austin Wagner
Software engineer · MCP & agent infrastructure · Irvine, California
About me
-
Backend and ML infrastructure: MCP servers and agent tool-routing, document ingestion pipelines, typed clients for undocumented APIs, and evaluation harnesses for machine-learning experiments. TypeScript, Python, Rust, PostgreSQL.
Tools and dashboards
-
Public records search
Find agendas, minutes, and staff reports from California public agencies.
In development -
Corvis
My research and operations platform for tools, knowledge, and AI agents.
Sign-in required -
Irvine City Council
City of Irvine meetings, agendas, minutes, videos, and council member profiles
-
OC City Councils
Contact info and public comment details for all 34 Orange County city councils
34 cities -
ASUCI Senate
UC Irvine student government meetings, attendance, and voting records
-
Metabolomics Analysis
Tree metabolomics data filtering and analysis report
What I work on
Research and engineering write-ups, grouped by topic. Open a group to explore.
Machine learning research 2
-
Measuring what a corpus installs
A seven-arm ablation on two published corpus-design interventions, run against real cited prose. One survived, one was inert under every condition its source specifies. Three of five results are nulls, and they are the reason the other two mean anything.
-
Composing one model out of several
A serving design assumed several separately trained compartments could be wired into one request. Measuring it first put the naive limit at two, and the run's own controls caught a result that read the right way and was not. A composition-aware recipe then moved four-compartment retention from −45.4% to +44.6%.
Reproducibility and research infrastructure 4
-
Reproducible GPU inference
One model and one set of weights produced a different answer on every NVIDIA card I ran it on. Owning the arithmetic order instead of the vendor library's made the output byte-identical across all nine: five architecture generations, two operating systems, one SHA-256.
-
Scoring a model you do not ship
Three ordinary training decisions compose into publishing a test score for a model that never leaves the machine. The manifest carried a validation loss from epoch 17 beside a test loss from epoch 20, neither labelled, and nothing errored. Measured, fixed, and confirmed on the cluster with both branches exercised.
-
Research cluster infrastructure
Eight research projects run on a shared university GPU cluster through one tool that will not let a job be built if it would fail, and can answer afterwards which job produced which artifact. Preflight that cannot be skipped, a budget the cluster itself does not enforce, and three failure conditions that look like health.
-
Research wikis that audit themselves
My seventeen research wikis share one audit engine, and each declares which kind of proof its citations owe: a PDF hash, a captured-page hash, a resolving vault query, or a git blob. A claim that cannot be checked that way fails the write, in the transaction and in the tool an agent calls.
Reverse engineering and system identification 2
-
A model that checks itself
I reverse-engineered the private protocol of an online game that publishes no documentation, then built an instrument that keeps proving the model is still right. It has since caught three of its own conclusions being wrong, including the ending of its best session: 84 kills, 0 deaths, 149 minutes.
-
Inferring an engine from the inside
A game ships as an obfuscated Java binary whose class names change between releases. I inject code into its running JVM to read real state and issue real orders, with no pixels and no synthetic input, and I made runs bit-identical per seed so a better strategy can be told apart from a lucky map.
Coordinating AI sessions 2
-
A board that refuses
Many AI sessions coordinate through one append-only board, and the board refuses what it cannot trust: a label that is not yours, a closure with no result, a delivery claim with no commit, a session id one keystroke from a real one. A surface that records is a log. One that refuses is infrastructure.
-
A page inside a proxy
A proxy that had only ever spoken to machines now renders a page for a person, on its own hostname, and every safety question it answers has exactly one implementation. The question that briefly had two answers cost a 403 naming no cause, and the fix is a guard that is itself guarded against gaining an unchecked fourth input.
Architecture and open source 4
-
MCP & agent infrastructure
Thirteen services and twenty-two shared libraries in one repository, held together by 47 architecture rules that fail the build, so my design decisions are enforced by CI rather than remembered in review. I wrote them across Python and Rust, because an invariant that holds in one language is not one.
-
Covenant Radar
I built a risk-prediction service for loan covenants, then generalised it so a new subject is a registration rather than a fork. Two more domains have since been added without editing the shared training, explainability or streaming machinery at all. Twelve model backends behind one interface, trained on 40 real public datasets.
-
ClearGBM
I wrote gradient boosting from scratch in numpy, ported it to a Rust core, and then deleted the numpy version, roughly 10k lines, because two implementations can disagree. The first one was the oracle I checked the port against; scaffolding comes down.
-
LightGBM
Nine merged pull requests: Python type annotations, a POSIX socket-timeout fix, and test coverage. Nine of twelve attempted, with the three that did not land in the distributed-training path.