fred67

fred67 is a group-chat AI that knows when to jump in, and when to stay quiet, while building living profiles of each member and the overall group’s vibe.

Comment

Inspiration

Group chats are noisy. Most bots make it worse—talking too much, at the wrong times, with zero sense of vibe. We wanted an AI that behaves like a considerate teammate: quick when helpful, silent when not, and actually learns the people in the room.

What it does

  • Selective speaking: Decides when to chime in (mentions, direct questions, confusion, milestones, de-escalation) and when to stay quiet (side-banter, solved threads).
  • Micro-replies: Keeps outputs ≤3 sentences unless asked for depth; playful, weird, never mean.
  • Memory: Stores lightweight facts about each participant and rolling notes about group norms (etiquette, inside jokes, boundaries).
  • Context steering: Grounds replies in the last window of chat to stay on-thread.
  • Safety rails: No hallucinated specifics, avoids stereotypes, and respects cooldowns to prevent spam.

How we built it

  • Frontend: Next.js (App Router), sticky composer, scroll-locked message list, presence pills.

  • Backend:

    • Supabase for auth (Google OAuth), Realtime messages, and storage of chat/memory.
    • Decision route (/api/decision): prompts an LLM for a STRICT-JSON policy verdict {speak, why, topic}.
    • Reply route (/api/fred): if and only if speak=true, crafts the actual message using windowed context.

  • LLMs:

    • JanitorAI completions for chat generation (OpenAI-compatible endpoint).
    • Letta for agent memory blocks (per-user “facts” and a group_dynamic block) and the decision JSON.

  • Guardrails: freshness gate (≤15s) so Fred only replies to recent human messages; last-speaker check to avoid back-to-back Freds; once-per-thread nudge cooldown.

Challenges we ran into

  • Edge vs server nuances: accessing req.url/origins and environment vars on Vercel vs local.
  • SSE/stream handling: stitching data: { "choices":[{"delta":...}] } into one coherent string.
  • Persona drift: keeping Fred playful but not mean; tightening the system prompt and adding “Do/Don’t” tables.
  • Double-posting: race conditions between realtime inserts and decision calls—fixed with freshness and “last is Fred” checks.
  • Env setup: mismatched keys (service role vs anon) and missing base URLs causing 401s/“too_old” no-ops.

Accomplishments that we're proud of

  • A bot that actually knows when to shut up.
  • Clean STRICT-JSON decision contract powering consistent behavior.
  • Live group memory that accumulates norms without leaking private info.
  • A lightweight, pleasant UI with sticky input, smooth scroll, and presence.

What we learned

  • “When to speak” is as important as “what to say.” Policy+JSON beats pure prompting.
  • Tiny fundamentals—cooldowns, last-speaker checks, and recency filters—dramatically improve perceived intelligence.
  • Memories need scope (per-user vs group) and limits (trimmed, summarized) to stay useful.

What's next for fred67

  • Multi-room & threads: per-channel norms, thread-aware decisions.
  • Memory UI: view/edit personal notes and group dynamic logs.
  • Better retrieval: embeddings + summaries for long-term context.
  • Bridges: Slack/Discord/Telegram connectors.
  • Moderation & safety: toxicity filters, escalation patterns, and red-team prompts.
  • Analytics: talk/silence ratios, helpfulness reactions, configurable guardrails.
  • Mobile PWA & notifications: light, fast, installable client.

Built With

  • jllm
  • letta
  • nextjs
  • oauth
  • supabase
Share this project:

Updates