HuddleX

Your Personal AI Council

HuddleX turns the experts and creators you follow into a personal AI council. Instead of asking one general-purpose assistant for one flattened answer, users can bring together a product thinker, technical expert, investor, creator, and contrarian voice around the same question.

The product is designed for moments when a decision benefits from perspective diversity: strategy, career choices, product direction, technical architecture, or creative positioning.

Each AI persona has its own knowledge base, communication style, voice, and thinking model. Users can switch perspectives inside the same conversation while keeping a shared thread memory, so one expert can build on what another expert already surfaced.

The prototype was built for the Always-On AI Coworker Hackathon at Boston Tech Week 2026, using the event scaffold for long-running conversational agents with memory, voice, and Rasa-based orchestration.

Conversation Flow

HuddleX structures a conversation around perspective switching. A user asks one question, chooses the persona they want to hear from, receives a source-aware answer, then invites another expert mind into the same thread.

Tech Stack & Architecture

Anti-Hallucination Design

• Facts, numbers, dates, and quotes are constrained to retrieved sources.
• ChromaDB stores expert knowledge from X posts, Wikipedia, essays, transcripts, and other public materials.
• A relevance gate decides whether a persona can answer from retrieved knowledge or should stay in framework-only mode.
• Responses expose grounding state so users know when an answer is sourced versus persona-style reasoning.

Memory Model

• Follow-up questions are rewritten into self-contained searches for better retrieval.
• Recent turns stay verbatim in a sliding window for conversational coherence.
• Older turns compress into a running summary to keep voice interaction fast.
• Threads persist as JSON, including which persona answered and which chunks backed the response.

Demo Video

My Contribution

• Developed backend voice-agent logic across FastAPI endpoints, Rasa CALM flows, and custom action-server behavior.
• Integrated the voice pipeline from user speech to ASR, Rasa/persona routing, LLM response generation, TTS synthesis, and voice reply.
• Implemented persona-agent backend surfaces for switching experts, updating user preferences, routing general chat, and calling persona RAG.
• Worked on grounding and memory behavior so voice interactions could use retrieved persona knowledge, recent history, and persistent thread state.

Hackathon Fit

• Built around persistent, long-term coworker behavior rather than a short one-off AI demo.
• Uses Rasa CALM and an action server for conversational orchestration.
• Combines voice input/output with persona-specific retrieval and shared memory.
• Targets resilient long-context behavior by pairing RAG, query rewriting, and bounded history.

Acknowledgements

Great building alongside the HuddleX team: @Jianxiang Ling, @Weihong Lin, and @Bolin lin.