Thank you for your interest in contributing to Sanket — AI-Assisted Indian Sign Language Accessibility System.
Sanket is an academic and research-oriented AI engineering project. To preserve scientific rigor, reproducible experimentation, and architectural modularity, all contributors and automated assistants must adhere to the engineering guidelines below.
- Keep the project modular: Separate vision ingestion, feature extraction, model inference, and language rendering into clean interfaces.
- Keep research decisions documented: Every significant algorithmic or architectural choice must be recorded in an Architecture Decision Record (
docs/decisions/). - Do not silently change technical decisions: Major shifts in models, dataset choices, or dependencies require discussion and documentation.
- Record experiments and results: Every ML experiment must be logged in
docs/experiments/with reproducible hyperparameters, metrics, and error analyses. - Keep GitHub Issues focused: Issues must correspond to concrete, actionable research or engineering tasks.
- Use Pull Requests for meaningful changes: Never push directly to
main. - Small, reviewable PRs: Prefer focused PRs addressing a single issue over massive monolithic submissions.
- Do not overclaim: Never claim "full continuous sign language translation" when the system currently implements constrained isolated sign recognition.
- Strict dataset licensing: Audit dataset licenses and redistribution rights in
docs/datasets/DATASET_REGISTER.mdprior to ingestion. - Privacy first: Never commit camera footage, raw biometric recordings, credentials, or private datasets.
The main branch is protected and represents validated, passing milestones. Direct commits to main are strictly forbidden.
main
↓
feature / research / fix / docs / refactor branch
↓
Pull Request
↓
Automated CI checks + Project Owner Review
↓
Merge into main
Use standardized, issue-traceable branch names:
feat/SANKET-XXX-description(orfeature/SANKET-XXX-...) - New features or pipeline additionsfix/SANKET-XXX-description- Bug fixes or defect resolutionsresearch/SANKET-XXX-description- Exploration, dataset surveys, feasibility investigationsdocs/SANKET-XXX-description- Documentation, ADRs, logs, or Wiki updatesrefactor/SANKET-XXX-description- Code improvements with identical external behaviortest/SANKET-XXX-description- Test suites, benchmarks, evaluation harnessessecurity/SANKET-XXX-description- Security hardening and vulnerability fixeschore/SANKET-XXX-description- Tooling, dependency, or configuration maintenance
Every Pull Request submitted to main must pass automated PR metadata verification (.github/workflows/pr-metadata.yml):
- Linked GitHub Issue: The PR description must explicitly include a closing reference (e.g.
Closes #33). - Milestone Assigned: Must be assigned to the active milestone (e.g.
Milestone 1 - Data & Isolated Sign Recognition). - Four-Dimensional Labels: Must carry at least one label from each required category (
type:*,area:*,priority:*,phase:*). - Conventional Commits PR Title: Title must follow standard formatting:
<type>(<scope>): <description> (SANKET-XXX). - Testing & Acceptance Criteria: Must include executed test proof and verified acceptance criteria checkboxes.
- All Pull Requests must be reviewed and merged by the project owner.
- Contributors and automated tools must create the branch, implement changes, and submit a detailed Pull Request.
- Never independently merge Pull Requests. The project owner controls progression.
- Work should proceed one issue at a time:
Issue → Implementation → PR → Human Review & Merge → Next Issue
Apply labels consistently across four standardized dimensions:
type:feature— New feature or pipeline capabilitytype:bug— Defect, regression, or calculation errortype:research— Investigation, literature review, dataset explorationtype:experiment— Machine learning experiment or baseline comparisontype:documentation— Documentation or record updatestype:refactor— Code refactoring without behavior changetype:chore— Repository maintenance, CI, dependencies, or configuration
area:dataset— Datasets, ingestion, annotations, and registersarea:vision— Computer vision, landmarks, frame extraction, MediaPipearea:ml— Machine learning models, features, training, inferencearea:temporal— Temporal modeling, sequential processing, LSTM/GRUarea:nlp— Natural language processing, sentence formation, grammararea:ui— User interface, feedback display, visualizationarea:api— API endpoints, service interfacesarea:deployment— Packaging, containerization, edge inferencearea:security— Security, privacy, credentials, compliancearea:testing— Unit tests, integration tests, benchmarks, evaluation
priority:P0— Critical blocker; address immediatelypriority:P1— High priority; necessary for current milestonepriority:P2— Medium priority; scheduled standard sequencepriority:P3— Low priority; nice to have or deferred
phase:0throughphase:6matching the roadmap phases
Follow the Conventional Commits specification:
<type>(<scope>): <short description>
[optional body]
[optional footer(s)]
Common types: feat, fix, docs, test, refactor, chore, research.
Example: feat(vision): add landmark normalization utility (Closes #1)
Before opening a Pull Request:
- Ensure all unit tests pass:
pytest - Ensure code style is formatted:
ruff check . - Verify no secrets, camera videos, or raw datasets are staged.
- Update corresponding documentation (ADRs, roadmap, project log).