Software that ships.
AI you can inspect.

Why I built it. Failing builds are a constant tax on teams: a dependency bump breaks a test, or a refactor misses an edge case. Greenlight handles the first pass—it reproduces the failure, finds a fix, proves it with the project’s own tests, and hands back a reviewable diff.

A hand-written loop. I didn’t use an SDK’s automatic function calling, so every tool call is visible, guarded, logged and counted against a budget. One provider interface has Claude and Gemini adapters, each keeping its native message history so thinking blocks and thought signatures survive between turns.

Guardrails by design. There is no general shell tool—the model can only run your test command. Test files are read-only by default, secret files are refused by every tool, credentials are scrubbed from the test environment, key-shaped strings are redacted, and each run stops at a hard step, token and cost limit.

Where teams work. A CLI, a web UI with recorded replays you can watch without an API key, and a GitHub Action that turns a failing CI run into a pull request for human review—it suggests, it doesn’t merge.

Measured honestly. On a 10-task synthetic suite scored with hidden tests, a small model (gemini-3.1-flash-lite) solved 10/10. On a harder SWE-bench-style benchmark of 22 real bug fixes from 8 open-source libraries—with the scorer verified by an oracle on all 22—it solved 4 of the 13 tasks it attempted (31%); free-tier quota stopped the other 9. Every fix’s diff is published.

Why I built it. GraphBench started as a correctness-first comparison of NetworkX and LadybugDB. It evolved into a tool that helps people interpret software dependency graphs and review architectural risk.

The workflow. Choose the guided sample or upload a CSV, confirm which component depends on which, then validate the graph before analysis. Reports turn graph structure into plain-language evidence and recommended next actions.

The analysis. PageRank highlights influential dependencies, betweenness centrality surfaces path bottlenecks, cycle detection exposes circular relationships, and K-core finds deeply embedded structural layers. The Graph Explorer offers four visual lenses with zooming, panning and focused subgraphs.

Failure impact. Select a component to inspect direct and indirect dependents, blast-radius percentage, impact depth and shortest dependency paths. These are structural review signals to interpret alongside runtime context.

Engineering foundations. CSV health checks and exportable JSON/CSV reports sit alongside the original latency, throughput, ingestion and resource benchmarks. Separate benchmark-review runs reached 1 million nodes / 5 million edges; that is a research scale, not a claim about the live analyzer’s capacity.

Why I built it. Getting comfortable in a new codebase often means jumping between files just to understand one request. I built a deployed RAG application that helps developers ask focused questions about public GitHub and local repositories.

From code to context. Tree-sitter parses six languages—Python, JavaScript, TypeScript, Java, C++ and Go—while preserving function, class and file metadata. Semantic search, BM25 and symbol retrieval find relevant chunks, then a cross-encoder reranks them before Gemini generates an answer.

Four ways to investigate. The Architect explains structure and execution flow; the Code Reviewer examines quality, security and maintainability; the Bug Finder investigates edge cases and failure conditions; and the Documentation agent explains APIs, classes and functions.

Follow the evidence. An interactive source explorer shows retrieved code, file paths and line ranges, with citation checks to connect answers to their source. A one-click sample repository lets visitors try the workflow.

What I measured. In an initial eight-question retrieval evaluation on the itsdangerous Python repository, the system achieved 87.5% Hit@5 and 0.750 mean reciprocal rank. These are preliminary results from one small evaluation.

Idea. Anomalies are often isolated with fewer random splits than typical observations.

Built. Tree construction, training and anomaly scoring from the original algorithm, connected to a FastAPI backend and Next.js dashboard.

Product. Upload data, run the model and explore the resulting anomaly visualization in one workflow.