All work

Problem. A change to a prompt, a model or a document set can make an assistant’s answers worse, and nobody notices until a customer does.

How it solves it. It grades answers against a document set with text rules, retrieval checks, citation checks and faithfulness rules, compares the run to a saved baseline, and exits non-zero on any regression, including a weakened test suite. It refuses to record a baseline from a failing run or to accept an empty one. A mutation suite plants defects and confirms each one is caught.

Who it helps. Teams that ship retrieval-based or support assistants and need a release check that fails loudly.

Technical. Python. Runs offline against a deterministic provider, with an optional OpenAI provider. CI on every push. MIT license.

Problem. Agents that call outside tools hang, retry forever, spend without limit, or run an unsafe action.

How it solves it. Every call gets a real deadline. Failures are sorted by type, and only temporary ones are retried. Cost and step ceilings are on by default, and the price of the next call is checked before it is made. Unsafe arguments are blocked before a tool runs, and every step is traced. A recorded run shows a real hang cut off by the watchdog and a halt at the cost ceiling.

Who it helps. Teams putting agents in front of real tools and real budgets.

Technical. Python. Runs offline with no keys. JSONL traces and an HTML trace report. MIT license.

Problem. When one model provider goes down or rate-limits, every app that depends on it stops, and repeated identical calls cost money.

How it solves it. An OpenAI-compatible proxy routes each request by rule, fails over across providers behind circuit breakers, applies rate limits, and serves repeated identical requests from an exact-match cache. A separate tracing package records tokens, latency and cost for each call.

Who it helps. Teams that call more than one model provider.

Technical. Python. Includes an HNSW vector index written without a library, document ingest and agent packages. Docker Compose. MIT license.

Problem. Support chatbots state things the help center never says.

How it solves it. It ranks a sample company’s help articles in the browser, shows the passage it used, and hands off to a person with a summary. An optional Claude step writes answers from those passages only, and its answer is rejected if it cites a passage it was not given or contains a number, email, phone or link that is not in the passages. The model step is off by default and has a daily cap.

Who it helps. Small support teams that want quick answers without invented ones.

Technical. JavaScript in the browser, BM25 retrieval, a serverless endpoint that accepts passage ids only. Live demo at support.dbhavery.ai.

Problem. Checking security controls against policies, configs and logs is slow, and a model’s guess is not an audit.

How it solves it. Each control gets a rule-based verdict with the reason and the matched excerpt. Secrets and personal data are masked, and the tool writes an HTML report and an append-only audit log. An optional model may add a reviewer note but cannot change a verdict.

Who it helps. Compliance reviewers and the engineers who prepare evidence for them.

Technical. Python. Runs offline. A leak check refuses to pass when it scanned no files.

Problem. Making a printable part normally takes CAD skills.

How it solves it. A plain-English description goes to a model that writes OpenSCAD. Compile, export and optional local vision checks send failures back to the model to repair.

Who it helps. Makers who need a simple part without learning CAD.

Technical. Python. Local by default through Ollama; Claude and Codex backends are optional.

Problem. Weather, warnings, quakes, fires, aircraft, ships and satellites live in separate apps.

How it solves it. It shows all of them on one 3D globe. Keyless public feeds run inside the app, so there is no account and no server to pay for.

Who it helps. Anyone who tracks weather and events in one view.

Technical. TypeScript, Cesium. Android, web and desktop.

Problem. Avatar quality is usually judged against a generic standard, which does not match what the task needs.

How it solves it. The paper judges photoreal avatars against the reference the task implies and sets out the identity-contract design that follows from it.

Who it helps. Teams building and evaluating AI-generated people.

Technical. Archived on Zenodo, DOI 10.5281/zenodo.21138188.

Problem. An on-site fuel-delivery service needs separate screens for customers, drivers, franchise owners and headquarters.

How it solves it. A static marketing site and a clickable demo of each role’s screens. All demo data is sample data.

Who it helps. Founders and investors evaluating the service design.

Technical. Static site. Demo at stafull.com.

Problem. A desktop assistant that can act on a computer needs a place where every action is checked and recorded.

How it solves it. Each action passes a policy gate that returns allow, ask, deny or draft-only, and every decision goes to a hash-chained, signed audit log. The Rust policy engine is built; the older Python voice tree is not yet routed through the gate.

Who it helps. People who want an assistant that runs on their computer and asks before it acts.

Technical. Rust and Python. Local models through Ollama.