AI Agents & Conversational Interfaces

AgentFloor visualizes Claude Code agents as pixel RPG characters executing tasks in a real-time office dashboard, solving the problem of monitoring multi-agent workflows at a glance. Built with Flask SSE streaming, vanilla Canvas rendering, and Python stdlib hooks that intercept agent lifecycle events, it displays live terminal output and agent state without external dependencies. Deployed on a VPS and wired to Claude Code via hook scripts, the system currently tracks concurrent agent activity with sub-second latency updates.

BarPhoneAgent automates B2B sales outreach for bar owners by scraping Yelp leads, qualifying them against competitor databases, and sending personalized SMS via Twilio with Claude-powered reply handling. Built with Python, SQLite, and multi-API orchestration (Yelp, SerpAPI, Twilio, Anthropic), it processes location-based lead batches with built-in rate limiting and dry-run safety. Successfully deployed with live SMS campaigns reaching 50+ qualified leads per location.

**Jarvis** is a screen-aware voice AI assistant for macOS that combines real-time speech-to-text (Deepgram), Claude vision/conversation models, and 18 desktop control tools into a unified Pipecat pipeline. It monitors the active screen, filters sensitive data (passwords, 2FA), and responds conversationally while executing system tasks—deployed as a persistent LaunchAgent with wake-word detection and self-modifying capabilities. Built with Python 3.13, Anthropic APIs, and ElevenLabs TTS; runs continuously at login with comprehensive logging.

**MayaSA** routes SMS and iMessage queries to Claude, which delivers hyperlocal San Antonio business recommendations enriched with real-time weather, event calendars, and persistent user memory. Built on FastAPI with Supabase persistence, it includes a loyalty system, business-owner portal, and automated weekly outreach—deployed on a VPS with live production traffic. The architecture prioritizes low-latency responses and conversation continuity across multiple messaging channels.

ResearchAgent decomposes complex research queries into atomic sub-questions, searches 8+ sources in parallel (Exa, Tavily, Brave, arXiv, Reddit, YouTube, etc.), and synthesizes credibility-scored reports using Claude with extended thinking. Handles business intelligence, academic research, and social sentiment at scale via async orchestration—outputs structured JSON reports with gap analysis and timelines.

Thirstistant automates mail and email triage for a solo user by extracting structured data via dual-LLM orchestration (Gemini Flash → Claude Sonnet), then routing extracted facts through iMessage approval gates to Paperless-ngx and calendar systems. The security model enforces strict separation: untrusted content stays quarantined in the reader LLM, credentials never reach any AI, and auto-login runs sandboxed behind biometric gates—eliminating the trust debt of most personal AI agents. Deployed self-hosted with rate-limiting and approval fatigue mitigations.