Rob has shipped software through a paid AI engineering engagement and built AI applications he operates himself.
At Genius Sports, he worked from ambiguous product requirements to shipped software as a part-time contract Forward Deployed AI Engineer. He used coding agents in structured iterations and built reusable practices for implementation, automated code review, and workflow handoff. He also architected the frontend's Storybook design system using Tailwind CSS and Radix.
That work connects agent-assisted development to the interface a team needs to maintain and ship.
In his own working environment at Clear Labs, the company he co-founded, Rob built a persistent agent platform. An orchestrator coordinates specialized Claude and GPT agents with shared skills, semantic memory, and messages between components. Queues and persistent sessions carry work across steps. Status reporting and restarts support its operation.
He uses it for research, implementation, and review, with evaluations and his own feedback around the work. His personal job-search system makes one human decision explicit: it remembers confirmed answers and tracks application state, but keeps preparing an application separate from approved submission.
Askrob, his public conversational portfolio, makes the model integration visible. A visitor asks about his work and receives a streamed answer grounded in a small collection of approved sources. Rob keeps those facts separate from voice examples and supplies the full approved collection on each turn.
The application also defines what a reply can do. The model produces conversational text. The application validates requests and handles actions around the conversation. A generated scheduling link does not create an appointment, and a contact follow-up requires explicit consent. The model workers cannot access his personal files or other agents' tools.
Rob owns the interface and the boundaries behind it.
That boundary continues through operation. Rob put model providers behind one interface for streaming, cancellation, readiness, and shutdown. The application can switch providers after a readiness check. A reply stays with the provider that started it, so a change does not move an unfinished generation elsewhere.
The application streams text as it arrives. Cancellation reaches the generation process, generation has a time limit, and concurrent work is bounded. Failures are recorded separately from completed answers. Rob records time to first text and total response time so he can investigate what a visitor experienced.
These are implemented controls. Provider switching is an explicit operation, with no automatic failover. How does the provider boundary work?
Rob also built a review path for what the application cannot answer. Askrob groups unanswered questions for review, and he can approve source updates for later turns. A visitor's message cannot publish a new fact about him or trigger model training.
Automated checks exercise knowledge approval, request validation, streaming, provider switching, follow-up consent, and rejected tool activity. Those checks verify application behavior. They do not make every generated answer correct.
What would Rob need to know before scoping our first integration?
Dema, the design system behind this site
