Taste Time Machine

Turn a half-remembered lyric, film, place, or family story into a cultural world worth sharing with Qloo's Taste Graph.

Comment

Inspiration

More than 55 million people worldwide experience cognitive decline, and millions of families struggle with a quieter daily problem: losing access to the cultural memories that shaped their loved ones.

Traditional search engines demand exact keywords they fail completely when someone only remembers "that slow accordion song from our wedding" or "a 1968 snowy train robbery western." On the other hand, raw generative AI models often hallucinate plausible-sounding fiction when presented with incomplete memories.

Taste Time Machine was inspired by reminiscence therapy and intergenerational storytelling. We set out to give older adults, family members, and caregivers a voice-first cognitive tool: start with whatever fragment you have a misheard lyric, a blurry visual setting, or an era vibe and ground it in verified cultural history rather than guessing.

What It Does

Taste Time Machine is a voice-first cultural memory investigation agent:

  • Voice-First Exploration: Speak or type an imperfect memory fragment into the Voice Explorer. Browser speech recognition transcribes words live, accompanied by an interactive WebGPU audio bloop.
  • Phonetic & Semantic Cleanup: A fast reasoning agent strips conversational noise, fixes common speech-to-text typos (e.g., misheard lyrics or names), and extracts key cultural anchors (medium, era, location, sound profile).
  • The Cultural Constellation (Qloo-Powered): Instead of returning a dry search result or an irrelevant corporate brand, the agent uses Qloo's taste graph to reconstruct a multi-domain world around the recovered memory (music, classic cinema, television, and historic dining/venues).
  • Interactive Signal Inspector: Every recovery displays a verified affinity indicator, giving users and judges transparent proof of how Qloo's taste graph correlated the lead discovery to surrounding cultural works.
  • One-Click Case Studies: Built-in scenarios allow reviewers and judges to test caregiver, family-film, and regional Motown queries instantly without a microphone.
  • Found Memories & Keepsakes: Confirmed discoveries are saved to a persistent browser archive and can be exported as a print-ready family keepsake card.

How We Built It & Qloo's Core Role

The application is built with a dual-engine architecture:

  1. Reasoning & Voice Agent:

    • Frontend: Next.js (App Router), TypeScript, Tailwind CSS, and Web Speech API.
    • Audio Visualizer: A custom WebGPU-rendered shader bloop with graceful fallbacks for low-power devices.
    • Agent Layer: High-speed LLM inference (OpenAI-compatible / Groq / NVIDIA NIM) to clean noisy transcripts, resolve ambiguous lyrics, and formulate structured search entities.

  2. Qloo Cultural Intelligence Layer:

    • Qloo’s API acts as the cultural grounding backbone. The agent queries Qloo across 250M+ cultural entities, filtering out non-cultural corporate entities and mapping multi-domain affinities.
    • The LLM reasons over the incomplete human clue; Qloo validates the cross-domain affinity so that recommendations reflect genuine shared audience tastes rather than statistical language guesses.

Challenges We Overcame

  • Real-Time Speech Inaccuracies: Spoken names and titles are easily distorted by microphone noise and accents. We engineered the agent to treat voice transcripts as noisy hints, repairing phonetic errors before running taste-graph queries.
  • Ambiguous Lyrical Clues: Many people only remember a hook or a misheard line. We designed the prompt pipeline to reverse-engineer lyrics into verified song titles rather than parroting raw conversational text.
  • Multi-Domain Latency: Sequential API requests across music, film, and dining can cause unacceptable UI delays. We implemented concurrent asynchronous calls (Promise.allSettled) with bounded timeouts to ensure the interface responds in under 2 seconds.

What We Learned

We learned that Qloo’s taste graph is uniquely suited for assistive cognitive technology. Traditional recommender systems are built for commerce and ads; using cultural taste as an associative retrieval prosthetic demonstrated how deeply interconnected human memories are across different media forms.

What's Next

  • Caregiver Companion Profiles: Enabling family members to pre-seed profiles with known childhood hometowns and family milestones.
  • Multilingual & Regional Expansion: Broadening vernacular support for regional dialects and global folk music.
  • Audio Streaming Integration: Connecting directly with Spotify and archive APIs so families can instantly hear the recovered song during caregiving sessions.

Built With

Share this project:

Updates

Submission history