Askrob: delivering an AI conversation inside a real product

A potential client can read a project page and still need one specific answer. What did I own? Why did I choose that architecture? What would I change for their product?

I designed and shipped Askrob so that conversation can happen inside the portfolio. It combines a streaming AI interface with approved source material, focused project conversations, scheduling and consent-based follow-up. I owned the product design, frontend, backend gateway and application logic around the model.

The first decision was to separate career facts from examples of my voice. An approved source catalog records where each document comes from, its topic and its priority. Private drafts stay outside the public collection.

The collection is small enough to load in full on every turn. That makes an approved correction available on the next question without maintaining a passage-retrieval index. The tradeoff is that context use grows with the collection.

A shareable project conversation opens with reviewed answers and suggested questions. Visitors can inspect the work, choose a question or write their own. The selected conversation gives the live chat context without turning a job description into evidence about my experience.

Actions belong to the application. A scheduling request opens the calendar flow. A follow-up requires a valid offer, contact information and explicit consent. Generated text cannot make a commitment on my behalf.

I built streaming responses, cancellation that reaches the generation process, time limits and bounds on concurrent work. Failed answers are recorded separately from completed ones. Time to first text and total response time help me investigate what a visitor experiences.

Model providers sit behind a shared interface. I can switch providers after a readiness check while an answer already in progress stays with the provider that started it. The interface does not need to be rebuilt around each provider.

Dema supplies the shared components and visual language. The live interface keeps native text entry and feature-specific question interactions, while the reference renders the same reusable components the site imports.

Tests cover approved knowledge, request validation, streaming, provider switching and follow-up consent. Unanswered questions become reviewable knowledge gaps. I can approve a source update for later conversations, rather than treating a visitor's message as a new fact.

The delivered product is live: a visitor can move from a project story to a specific answer and an explicit next step. Its source controls and application checks make that workflow reviewable. They do not guarantee every generated answer is correct.

If you have an AI feature to get into people's hands, I can own the interface and the application behavior around the model. We can scope the knowledge it needs, the actions it can take, and what happens when a response fails. Ask how I would approach your first integration.