stevens
Team led by a Stevens CS student experienced in Python ML pipelines, scikit-learn, and Java, specializing in NLP and end-to-end model development.
YouTube Video
Project Description
This project reimagines Claude not as a chat window, but as a controllable reasoning interface. Instead of hiding the model behind a linear message stream, the UI exposes structured controls—creativity sliders, reasoning modes (Fast, Analytical, Step-by-Step), output formatting, and editable final responses—turning the interface itself into the core innovation. Users don’t just “chat”; they steer cognition.
Note the audio for tarvus was not captured but it was reciting my week one plan😭😭.
Stability is achieved through a layered architecture: a Next.js frontend orchestrates requests, a Redis Agent Memory Server manages working and long-term memory via explicit /v1 endpoints, and local embedding models remove external dependency failures. The system gracefully handles memory extraction, session IDs, debounce scheduling, and fallback logic to prevent crashes or silent failures. Redis ensures session consistency, while local SentenceTransformers embeddings eliminate OpenAI key reliance and reduce runtime instability.
The interface shatters traditional chat conventions by separating reasoning from output, allowing users to refine, regenerate, and edit responses within the same memory context. The “Reasoning Summary” panel surfaces structured cognition without overwhelming users, and the Tavus Live Audio mode introduces a novel interaction pattern: users can shift from typed interaction to a real-time, voice-driven AI presence within the same stateful session. This transforms Claude from a static assistant into a dynamic, embodied interface.
Claude’s advanced reasoning is central to the generative experience: structured prompts, controlled output modes, and memory-aware responses allow the system to function as a decision engine rather than a conversational toy. The UI becomes an extension of the model’s reasoning process—transparent, adjustable, and persistent.
Technologies Used:
Next.js (React + TypeScript)
Tailwind CSS
Redis Stack (Docker)
Redis Agent Memory Server (v1 endpoints)
LiteLLM (provider abstraction)
Anthropic Claude (claude-sonnet-4-5)
SentenceTransformers (local/all-MiniLM-L6-v2) for embeddings
Tavus Live Audio API
Node.js API routes for orchestration
Together, these technologies enable a responsive, intelligence-driven interface where human-AI interaction is not constrained to chat—but evolves into structured control, persistent memory, and live conversational presence.
Prior Work
No prior work. This project was built during the hackathon.