Astra AI

At a glance

Scope

Groq agent with four server tools: résumé lookup, projects, Cal.com summary, page navigation.

Shipped

Tool loop behind ASTRA_AGENT_MODE; eval harness, multi-profile lenses, TTS/STT.

Problem

Visitors needed a faster way to explore the portfolio on mobile than reading every section in order—without bolting on a generic third-party chat widget.

Constraints

Replies must use only public resume and project data. Keep prompts small; model must retrieve facts via tools instead of inventing them. UI should feel native to this site.

Approach

Astra runs a server-side Groq tool loop: get_knowledge, get_project, get_calendar_summary, and navigate_to_section. The chat panel shows which tools ran, scrolls the page via ui_actions, and falls back to grounded copy when synthesis fails.

Outcome

Honest “applied AI agent” story for interviews—tool trace, eval cases, guardrails—not a black-box chatbot.

Signals

Scroll and Cal hooks cut “where do I click?” friction for recruiters on phones; tool trace makes grounding visible.

Timeline

• Frame

  Trust rules: public data only; clear errors when APIs are missing.

• Ground

  Resume + projects as structured JSON; tools retrieve slices on demand.

• Agent

  Groq tool loop, ui_actions, synthesis fallback, npm run astra:eval golden cases.

• Integrate

  Chat panel tool trace, voice strip, TTS/STT, section-aware suggestions.