Specter
by Flora Dixit, Aritro Bhattacharjee, Bhagirath Bandigari, Divyash Nath
Specter learns skills on your Mac, remembers them in GhostWiki, teaches back with Tavus PAL — publish caught skills on Skills Hub.
The idea
What's it about, and what makes it original? What problem or space does it explore? Specter is macOS desktop tutor that floats over real apps, watches how you work, saves workflows into local GhostWiki memory, and teaches you back with screen-aware walkthroughs and Tavus PAL video. Skills Hub lets you catch learned skills and share them — anyone can train with PAL. Original part: not static RAG or one-shot screen bot. Closed loop — observe, remember, teach, lint gaps, user corrects, re-query. Memory from your sessions, not generic docs. Walkthrough-first: you click, Specter guides. PAL gets real skill context. Problem: "How do I do this again?" Generic AI forgets your workflow. Docs go stale. Screen bots brittle. Space: personal computer-use memory + tutoring — bridge between session recording, procedural knowledge, and conversational AI on desktop.
What it does
Summon overlay anywhere (double-tap Shift). Chat or voice with Claude grounded in GhostWiki — ask what it remembers about past workflows, get answers with real sources. Request walkthrough ("walk me through creating an event") — Specter captures screen, resolves UI targets, shows guide ring where to click; you click, it advances. GhostWiki ingests session workflows as markdown wiki pages; query, lint for missing steps, correct, re-query — memory improves. Memory dashboard (Cmd+Shift+M) shows saved entries, corrections, skill progress. PAL back-and-forth: Click Tavus circle → live video conversation. Talk about skill you selected on Skills Hub or workflow on your Mac — event creation, calendar steps, anything GhostWiki captured. PAL gets context: moves learned, workflow steps, memory stores from past calls. You ask → PAL explains step-by-step, references your actual workflow data → you follow on-screen guide ring or ask "what's next?" → PAL adapts. Catch skill on Hub → publish → anyone trains with same PAL using that skill's content. Mac session auto-syncs dex to Hub — learned moves update after you work.
How it works
Two surfaces, one PAL. PALmaker (Tavus) — built Specter PAL: personality, perception (see/hear), memory_stores per user so PAL remembers you across calls. TAVUS_PERSONA_ID + replica from PALmaker. Server route POST /api/tavus-conversation hits Tavus POST /v2/conversations with skill body as conversational_context — never expose API key client-side. Lovable — shipped Skills Hub at live URL. React app: browse skills, battle/learn steps, catch to publish, TRAIN with PAL iframe. Vercel serverless API for journey sync, publish gate, Tavus conversation proxy. Mac pushes dex via POST /api/journey after GhostWiki compile. Magic Canvas (Tavus CVI) — interactive cards inside live PAL call (questions, pickers, charts). Specter pairs PAL voice/video with Lovable app shell for full skill journey; Mac overlay pushes live screen context mid-call via Daily append_llm_context so PAL sees what you're doing, not just static prompt. Mac layer (Specter Electron) — screen capture + multi-tier target resolver (Playwright → AX → vision). GhostWiki = local Python FastAPI sidecar (/ingest, /query, /lint) over markdown workflows. Claude for chat/planning, NVIDIA NIM for screen vision, Whisper for voice. Tavus PAL embeds in overlay circle via Daily iframe; local TTS pauses while PAL live. Flow: work on Mac → session compiles to GhostWiki → syncs to Hub → catch publishes skill → anyone opens Hub → PALmaker PAL teaches from skill context + Tavus memory → back-and-forth in iframe while Lovable UI tracks progress. macOS desktop tutor that floats over any app (double-tap Shift). Watches how you work, saves workflows to GhostWiki (local markdown memory), recalls them in chat, walks you through steps with on-screen guide ring. Tavus PAL = live video face when teaching. Skills Hub (Lovable) = public dex — browse skills, train with PAL, catch learned workflows to publish. Not static RAG — living manual for your software use.
What's next
Magic Canvas in-call steps — PAL surfaces each workflow step as interactive card inside Tavus video (confirm, pick target, mark done). Webhook feeds back to GhostWiki + Skills Hub dex. Teaching becomes visual back-and-forth, not voice-only. Shared skill marketplace — anyone catches workflow on Hub; others train with same PAL from published skill body. Mac learns locally, Hub publishes globally. Personal memory → shareable PAL-teachable skills at scale.
■ Event Recap · June 27, 2026 · Tavus HQ
PALmaker × LovableHACKATHON
Built with PAL Maker + Lovable, together with the ProductTank SF community — Tavus HQ, San Francisco, June 27, 2026.
Chapter 01 / 04
Doors open.
Welcome, builders — badges on, credits claimed, countdown running.
The winners
On the podium
■Judges' favorites
Ten more that stood out
The full lineup
A–Z
■ The judging panel
Eleven judges. Every build scored.
★ THE JUDGING PANEL
- 01Saswat MishraProduct @ Hyperbound
- 02Gagan BhatApplied AI @ Anthropic
- 03Gaurav MahajanEngineering @ Apple
- 04Ankit SultanaEngineering @ Notion
- 05Karsh PandeyProduct @ Virta Health
- 06Zane HomsiProduct @ Glean
- 07Kelly PengFounder @ First Intuition
- 08Rohan BenkarEngineering @ Coursera
- 09Darius KarelEngineering @ OpenAI
- 10William GaoMentor @ StartX · stealth founder
- 11Karthik Ragunath Ananda KumarAI Research @ Tavus
■With thanks
Our hosts
■June 27, 2026