I build evidence-driven tools for reliable AI, data quality, accessibility, developer workflows, and robotics.
- Reliable AI and evaluation: explicit evidence boundaries, reproducible experiments, and uncertainty-aware decisions
- Developer tools: local-first debugging, repair verification, test automation, and inspectable artifacts
- Data and annotation quality: deterministic checks with human review rather than unsupported model claims
- Accessible software: keyboard-path testing and regression evidence for real user journeys
- Physical AI and robotics: telemetry-based evaluation, ROS 2 safety components, and honest simulation/hardware boundaries
Experimental visual-evidence verification for annotation-quality and physical-action assessment. Release v0.2.8 records 217 Python tests and 19 Rust tests; the repository clearly limits its results to controlled prototype validation rather than independent physical or human-subject validation.
Python FastAPI computer vision Rust Playwright
A local-first Python debugging and repair-evidence lab. Its workflow separates a proposed or applied repair from a repair that has been independently verified, and documents approval, path-containment, secret-redaction, and optional container-verification boundaries. Release v0.2.1 and CI are published; the documented suite reports 86 passing tests.
Python FastAPI agent workflows Docker security controls
A controlled accessibility-testing proof of concept that compares equivalent mouse and keyboard booking journeys, reads saved state independently, and generates Playwright regression tests. Recorded verification covers 12 backend tests and 5 Chromium E2E tests; screen-reader, complete contrast, and cross-browser verification remain outside its tested scope.
TypeScript React Fastify Playwright accessibility
An annotation-quality workflow for CSV structure, bounding boxes, image dimensions, taxonomy consistency, reviewer decisions, and evidence export. The implementation uses deterministic checks and human review; it does not claim universal annotation correctness.
Python image QA human review pytest GitHub Actions
A DROID Episode 0 evaluation prototype that combines real recorded telemetry with separately reviewed visual evidence. Its reports distinguish measured episode signals from exploratory detector/tracker analysis and do not present the work as independent physical validation.
Python robotics telemetry computer vision reporting
ROS 2 Jazzy navigation and safety source with successful source/build CI. End-to-end Gazebo, Nav2, and real-hardware validation are explicitly documented as pending.
C++ ROS 2 Jazzy Nav2 Gazebo CI
Flisk Founding Contributor #10 of 100, Flisk Legend for 100+ merged pull requests, and Flisk Brown Belt in the 4–1 kyu category. The Flisk Hall of Fame records 101 merged pull requests, 101 issues fixed, and a highest individually confirmed rank of 4 kyu.
Languages: Python · TypeScript · JavaScript · C++ · SQL · Bash
Systems and frameworks: FastAPI · React · Node.js · Playwright · Streamlit · PostgreSQL · SQLite · ROS 2 · Docker · GitHub Actions
Working practices: pytest · CI/CD · reproducible evaluation · structured logging · human-in-the-loop review · evidence-oriented documentation
- Evaluation and observability for AI systems
- Trustworthy agent workflows and approval boundaries
- Dataset, annotation, and response-quality tooling
- Accessibility regression testing
- Computer vision and physical-AI evidence
- Open-source maintenance and contributor experience
- Evidence over claims: state what was tested, how it was measured, and what remains unknown.
- Reproducibility by default: keep tests, commands, fixtures, and artifacts inspectable.
- Human accountability: use automation to support—not obscure—review and responsibility.
- Secure boundaries: minimize data exposure and make risky operations explicit.
- Accessible systems: treat keyboard and assistive access as engineering requirements.
- Portfolio: kavixo.in
- LinkedIn: linkedin.com/in/putlurusandeep
- GitHub: @sunny-kavixo