Navdeep Saini

Ship PRSense v1.0. Deliverables: (1) rebuild the benchmark harness and rerun against current flagship models (OpenAI, Anthropic, Google) and local Ollama models via together.ai, publishing precision/recall and grounding metrics; (2) patch-based review mode for distributed git workflows (mailing-list patch series, stacked PRs); (3) raise test coverage on core review and indexing paths to a defensible bar. Approximately two months of full-time work.

Close the known gaps in the TypeScript symbol-graph pass: heritage relationships (implements / extends), declared-type structural references, and signal calibration to reduce the false-positive rate observed in dogfooding without regressing the false-negative profile. Includes the queued internal refactors (HttpsCloneRepositorySource consolidation, ContextProvider and Reporter contract tests). Approximately two months of full-time work.

Build a replay harness that runs PRSense over a curated corpus of historical merged pull requests from well-known open-source repositories, comparing signals against subsequent revert/hotfix/bugfix commits to derive an empirical precision-recall curve. Output: a public benchmark dataset and a calibration loop that catches signal-quality regressions before they ship. Approximately two months of full-time work.

Extend PRSense's depth-first (AST chunking) and breadth-first (cross-file symbol graph) review architecture beyond TypeScript. Deliverables: tree-sitter-backed AST chunkers for Python and Java; cross-file symbol-graph context providers for both languages; per-language calibration using the PR Replay harness; refreshed benchmark covering each language against flagship and local models. Approximately five months of full-time work.

Annual support covering security patches, dependency and SDK updates (Octokit, gitbeaker, Forgejo API, LLM provider SDKs), bug triage, and release management. Roughly one month of focused work spread across the year.

Any contribution toward continued development is appreciated.