SynapseTaste

Inspiration

Modern Large Language Models (LLMs) are exceptional at conversational reasoning, but they struggle with genuine human taste. When asked for recommendations—like pairing a vinyl record with a boutique dining spot or finding books matching a specific cinematic tone—they default to mainstream cliches and surface-level keyword matching.

We were inspired to build SynapseTaste to solve this grounding problem. We wanted an autonomous agent that understands the subtle, cross-domain connective tissue of culture: bridging semantic reasoning with empirical data from Qloo’s Taste Graph.

What it does

How we built it

We architected SynapseTaste as a multi-stage autonomous pipeline:

Agentic Orchestration: Built with Python and LangGraph to manage stateful, multi-hop queries and error recovery.

Cultural Knowledge Layer: Integrated Qloo's REST APIs as an empirical constraint layer. The agent queries cross-domain affinity matrices:where denotes the latent cultural embedding vector and provides normalized similarity scoring. User Interface: Developed a clean, responsive front end using Next.js and Tailwind CSS, featuring interactive graph-edge visualizations for discovered items.

Challenges we ran into

Entity Disambiguation: Translating colloquial conversational cues (e.g., "something like 90s shoegaze but with a warm coffeehouse feel") into canonical Qloo entity identifiers required robust fallback fuzzy-matching and schema-validation prompts.

Latency Optimization: Multi-agent chains querying external graph endpoints in real time introduced high response latency. We resolved this by implementing asynchronous batch querying and a lightweight in-memory cache for popular node neighborhoods.

Avoiding Overfitting: Initial models favored overly identical media clusters. We fine-tuned our selection loops to promote cross-category discovery without losing cultural relevance.

Accomplishments that we're proud of

True Multi-Hop Cultural Links: Successfully demonstrating high-precision recommendations between completely disjoint categories (e.g., French New Wave cinema to modern culinary styles).

Zero Hallucination Rate: Eliminating fabricated titles and nonexistent venues by forcing our reasoning agent to ground every terminal output in valid Qloo entity identifiers.

Intuitive UI: Delivering an interactive visualization that lets users trace the graph connections explaining why an item was curated for them.

What we learned

Cultural tastes operate on latent, non-linear vectors that pure language token prediction cannot reliably replicate on its own.

Hybrid workflows—pairing the fluid natural-language interface of an LLM with structured graph databases—outperform standalone prompting pipelines in factual and subjective tasks alike.

Balancing discovery (novelty) against familiar affinity is the critical differentiator between a boring recommendation list and an exciting user experience.

What's next for SynapseTaste

Collaborative Group Curation: Expanding the agent to synthesize intersecting graph profiles for multiple users (e.g., planning a movie night or dinner party that balances everyone's distinct taste).

Dynamic Context Ingestion: Incorporating real-time contextual factors such as local weather, time of day, and geographic location directly into the candidate ranking function.

Autonomous Tool-Use Actions: Enabling the agent to interact with downstream platform APIs (Spotify playlist creation, OpenTable bookings, Letterboxd sync) directly from the curated interface.

Built With

• agentic-ai
• artificial-intelligence
• async-await
• autonomous-agents
• fastapi
• git
• graph-database
• json
• langchain
• langgraph
• machine-learning
• next.js
• openai
• prompt-engineering
• python
• qloo-api
• react
• recommendation-system
• rest-api
• tailwindcss
• typescript
• vector-search