Software engineer in Arizona building and learning in public around AI-assisted software development, reliable agent workflows, local-first tools, and evidence-traceable research.
I care about systems that are useful and inspectable: clear boundaries, reproducible verification, and honest documentation of what works and what does not.
- Agent Harness — a personal agentic development harness that turns an Obsidian inbox ticket into a tested pull request with an explicit human review loop. Includes a live demo and end-to-end workflow.
- Cognitive OS — an experimental local-first control plane for turning ongoing intent into dependency-aware, reviewable work without silently granting execution permission.
- Neglected Science — an evidence-traceable metaresearch program for finding computationally testable questions and rejecting weak candidates before experimentation.
- Where coding agents become reliable enough to use as engineering infrastructure rather than demos.
- How software systems should represent intent, permissions, memory, verification, and failure when AI can take actions.
- How strong technical communities turn individual experimentation into shared engineering knowledge.
- How research and product ideas can be tested rigorously before large amounts of effort are committed to them.
- Emergency AI — an offline-capable engineering prototype exploring bounded emergency decision-support interfaces. It is explicitly not an emergency service or clinically validated system.
- Neon Stereo and other focused applications remain independent experiments rather than parts of one oversized platform story.
- Prefer runnable prototypes over slideware.
- Keep automation reviewable and interruptible.
- Separate demonstrated behavior from aspirational claims.
- Document failure modes and current limits alongside features.
- Use public projects to learn, test ideas, and have better technical conversations.
For the fuller project map, see docs/PORTFOLIO.md.
Contact: tarunsp23@gmail.com