Stephen Kamau

A computer-vision workflow covering object detection, multi-object tracking, trajectory analysis and instance segmentation, with persistent IDs and annotation validation across video frames.

• — Tracks continuity through motion and occlusion.
• — Uses Python and JSON checks to surface annotation defects.

A side-by-side model review interface for comparing two answers against one shared dataset, checking completion, factual accuracy, instruction following and unsupported claims before making a preference judgment.

• — Turns rubric criteria into visible, auditable review fields.
• — Shows reasoning quality alongside source data and issue findings.

A Python and JSON review toolkit that brings text prompts, visual inputs, structured records and tool interactions into one evaluation flow for quality scoring, pairwise comparison and structured error analysis.

• — Scores factuality, instruction following and grounding.
• — Supports structured error output and reusable review records.

A longitudinal compliance benchmark that tests multi-agent reasoning across a protocol PDF, SQLite records, telemetry and incident narratives before producing a strict adjudication output.

• — Seven coordinated subtasks separate evidence extraction from final synthesis.
• — Docker and executable verifiers support reproducible scoring.

A Flutter marketplace for university communities with Firebase services and Gemini-assisted image-to-listing workflows designed to reduce manual listing effort.

• — Turns a product image into structured listing assistance.
• — Combines app development with practical multimodal automation.

A lightweight terminal utility that presents available media formats, accepts a format choice and reports the selected download flow with clear runtime feedback.