◆ A retrieval-augmented mental health companion
Talk through it,
without judgment.
SafeSpace listens with empathy, detects emotion and intent, and grounds every response in evidence from professional counseling literature — so you always know where the help is coming from.
Not a replacement for therapy or crisis intervention. In emergencies dial 988 (US) or findahelpline.com
◆ What it does
Empathy, grounded in evidence.
Every reply is shaped by an NLP pipeline that understands you, then sourced from a vector index over 17k professional counseling pairs and a curated PDF library.
Emotion-aware
DistilBERT classifies emotional tone across fear, sadness, joy, anger and more — visible inline, never guessed.
Intent routing
Zero-shot intent classification routes greetings, questions, and crises through the right pathway.
Evidence-based sources
Hybrid BM25 + semantic retrieval cites the actual passages behind each response.
Crisis safety net
Keyword and signal detection surface 988, Crisis Text Line and chat resources the moment they're needed.
Multilingual detection
TF-IDF language ID across 20 languages so we always meet you in the language you're typing.
Context across turns
Symptoms, triggers, severity and duration accumulate as you talk — never re-explain yourself.
◆ Under the hood
A pipeline, not a black box.
From your first sentence to the final answer, every step is visible.
- 01
Language detection
TF-IDF + Logistic Regression across 20 languages.
- 02
Emotion classification
DistilBERT fine-tuned on 6k labelled messages.
- 03
Intent routing
Zero-shot via Groq LLM — greeting, gratitude, mental-health, out-of-scope.
- 04
NER + query rewrite
Extract symptoms, triggers; rewrite for retrieval.
- 05
Hybrid search
BM25 keywords + semantic similarity over the Qdrant index.
- 06
Grounded generation
LLM composes a response strictly from retrieved chunks, with citations.
◆ The team
Built with care.
Mohamed Emad Zaky
ML & Backend
Ziad Mahmoud Amr
ML & Backend
Tech · Python · PyTorch · Hugging Face · Qdrant · FastAPI · Groq