AURORA⁺ is a real-time emotional support chatbot designed to engage users with brief, empathetic conversations. It uses modern LLM infrastructure, sentiment-aware embedding verification, and personalized multimedia responses (GIFs, Spotify, Markdown, etc). Built for web, Jetson devices, and edge AI.
- 🧠 LLM-driven therapist (e.g.,
GroqAPI or OpenRouter → LLaMA 3 / Mistral 7B) - 🎵 Dynamic Spotify track embedding from chat context
- 😄 Emoji picker and GIF responses via GIPHY
- 💡 BERT-based sentiment verification layer
- 🔗 LangChain for modular prompt pipelines & memory
- ⚡ Optimized inference for Jetson Nano/Orin using quantized transformer models
- 🌗 Dark/light theme toggle + animated UI
- 🧩 Floating Cortana-style assistant UI (optional)
| Area | Tool / API |
|---|---|
| Frontend UI | React + Tailwind + Framer Motion |
| LLM Inference | Groq API / OpenRouter |
| NLP Tokenization | BERT Tokenizer via transformers |
| Prompt Engineering | LangChain + Custom Routing Layers |
| Embedding Verification | bert-base-uncased (HuggingFace) |
| GIFs & Emojis | Giphy API + EmojiPicker |
| Music Recommendations | Spotify Web API |
| Edge Deployment | Jetson Nano / Orin + ONNX Runtime |
USER INPUT
↓
[LangChain Router]
↳ Verify Emotion → BERT Tokenizer + Sentiment Model
↳ Enhance Prompt → Memory + Intent-aware Chains
↳ Select LLM → Groq / Mistral / LLaMA
↓
LLM RESPONSE
↳ Parse Markdown, Embed Spotify, Detect Image Markdown
↳ Search GIF via alt-text
↓
REACT FRONTEND
↳ Animate Bubble + Response
↳ Display Floating Chat UI
| Component | Description |
|---|---|
bert-base-uncased |
Verifies emotional tone of user messages (negative, positive, neutral) using HuggingFace pipelines (transformers) |
Groq + LangChain |
Hosts LLMs like LLaMA-3 and Mistral via super-low-latency inference with context-aware chains |
LangChain PromptNode |
Custom chain built with chatWithOpenRouter() wrapper + Spotify/Giphy hooks |
The project optionally runs on Jetson Nano/Orin using quantized ONNX versions of BERT & distilled transformer models.
Optimizations include:
onnxruntime-gpufor CUDA- Sentence embeddings via
MiniLM-L6 - Model quantization to
int8for edge performance
git clone https://github.com/yourusername/aurora-chat.git
cd aurora-chat
pnpm install
pnpm devOr run backend BERT/Groq in api/:
cd api/
pip install -r requirements.txt
uvicorn main:app --reloadGROQ_API_KEY=sk-...
SPOTIFY_CLIENT_ID=...
SPOTIFY_CLIENT_SECRET=...
GIPHY_API_KEY=...- 🔊 Speech-to-Text via WebRTC or Whisper
- 🧠 Memory threads & avatar stateful logic
- 📲 PWA (Progressive Web App) support
- 🌐 Voice-based chat via Jetson edge microphones
| Term | Meaning |
|---|---|
| LangChain | A framework to build modular and prompt-based LLM pipelines |
| BERT Tokenizer | Tokenizes input into word pieces for sentiment analysis |
| Groq | High-speed inference engine for LLMs like LLaMA 3, Mistral |
| Jetson AI | NVIDIA’s edge computing devices for real-time inference |
| Emotion Router | A logic layer that adjusts LLM prompts based on emotion classification |
Feel free to fork and extend!
DM @yourhandle for collabs or deployment help on Jetson boards.