agent-gateway
Admission control, budgets, checkpoints, and watchdogs for coding agents that run unattended, because an agent that has stopped working does not report the fact.
Result241 tests passed, 1 skipped
The production suite stays private because it is company work. The public repositories below show the same method, with the code open to inspection.
Agents and LLM systems
Systems that carry their own checks.
An agent that has stopped working does not say so, and a retrieval score does not explain itself, so each project below carries its own checks.
tm-rag-copilot
Retrieval over a synthetic translation memory with deterministic QA. The experiments record what failed as well as what worked, including a hybrid search that lost at three of four depths.
ResultRight match in the top 10 for 82.9% of queries with a language filter
docx-mcp
A read-only MCP server that lets an agent read a Word document as it would stand with every tracked change accepted.
Result67 tests passed; six tools, none of which can write
rosetta-bench and rosetta
A harness for comparing prompt formats across model providers, and the tooling that keeps one prompt in a human form and a machine form.
Result38 and 50 tests passed; no model benchmark results published yet
Research
Questions written down before the code.
The papers below are manuscripts, written to inform the repositories that accompany them. None has been submitted to a journal or peer-reviewed.
Pseudocode as a guardrail
A single-case study in which a language model drafts a pseudocode specification, a compiler enforces its contracts, and execution finds what review missed.
ResultEight defects found by execution and logged
Detecting LLM text watermarks
A research lab and evidence synthesis on what statistical watermarks can and cannot show. It measures and detects, and it ships no tooling for removing watermarks.
StatusEmpirical validation of the application is pending
Software
Applications that hold up outside a notebook.
The same habits carry over to desktop software and signal processing, where the record of each decision matters as much as the code.
BachelorPad+ Lite
A Rust text editor with Notepad's features and a local note layer that never touches the network.
Result1,104 test functions in the source; 94 recorded decisions
BachelorPad+
The larger Rust text-processing application for Windows and Linux.
Result2,026 test functions in the source; 68 recorded decisions
SlideBridge
Converts PowerPoint decks into review-ready Word documents with slide images and tables.
Result45 tests passed
Amplified Futures
C++ DSP modules for VCV Rack 2, built for drone, feedback, and detuned oscillator systems.
Method
Small experiments rather than large claims.
- 01QuestionEach project begins with a question and a written design.
- 02BuildCoding agents implement the design under test suites and review gates.
- 03MeasureEach change faces tests or an evaluation before it stays.
- 04RecordWhat fails goes into the record with its numbers.
The work proceeds through small experiments rather than large claims, because a change small enough to measure is also small enough to undo. Each one faces tests or an evaluation; what works stays, and what fails goes into the record with its numbers, since a negative result on paper saves more time than a success nobody measured.
However, a tested system still needs an argument behind it, so each project begins with a question and a written design. Coding agents then implement that design under test suites and review gates, which keeps the judgment with the author and the labor with the machine.