I build AI systems people can inspect and operate.
I am Petru Arakiss. I care about systems that make their evidence visible: retrieval with sources, agent workflows with boundaries, failure states people can understand, and writing that keeps claims tied to what can be checked.
background
I like software that shows its work. A useful AI system should make sources, permissions, uncertainty, costs, and failure states visible enough that an operator can make the next decision.
The public thread in my work is evidence: retrieval checks, evals, traces, policy boundaries, release gates, and product screens that explain what happened instead of asking people to trust a fluent answer.
Private production work stays bounded. Public repositories and essays are where the surrounding engineering ideas become visible enough to inspect.
Principles
Build for production from day one
Evidence before confidence
Documentation is part of the deliverable
Async by default
Design for failure first
Clarity over magic
Focus
Evidence before confidence
A confident answer is cheap. The useful question is what source, check, trace, or release gate supports it.
Boundaries before autonomy
Agents need clear tool boundaries, stopping conditions, approvals, and records before they deserve more freedom.
Product before theater
The interface should show the failure state, audit trail, escalation path, and next valid action without turning the user into a debugger.
Stack
- TypeScript
- Python
- Rust
- JavaScript
- SQL
- Retrieval-Augmented Generation
- Semantic Search
- Vector Databases
- Qdrant
- PostgreSQL/pgvector
- Evaluation loops
- React
- Next.js
- TypeScript
- Web Components
- Lit.js
routing
The CV carries role history. Projects carry inspectable code. Writing carries the reasoning behind the work.