PersonaX: Your Digital Reflection - A Locally Deployed, Biometric-Aware, Emotion-Conditioned Personalized AI Assistant
International Journal of Advanced Research in Science Communication and Technology · 12.06.2026 · 362
DOI: 10.48175/ijarsct-36539 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Research in Science Communication and Technology, online erschienen am 12.06.2026. Verfasst von Mr. T. Arivanantham und 3 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben.
Abstract (Originalsprache)
Existing conversational AI assistants share three fundamental limitations: they are cloud-dependent, stateless across sessions, and completely unaware of who is speaking or how that person feels. This paper presents PersonaX: Your Digital Reflection, a fully offline, biometric-aware, emotion-conditioned personalized desktop assistant that addresses all three limitations simultaneously on consumer CPU-only hardware. PersonaX is implemented as a six-layer pipeline: mul-timodal input (text and audio), face-based automatic user identification via InsightFace buffalo_l (512-dimensional embed-dings, cosine similarity threshold 0.45), real-time 7-class emotion detection via DeepFace gated on biometric confirma-tion, dual-memory retrieval combining MongoDB short-term context (6-turn bounded deque) with FAISS IndexFlatIP long-term semantic memory (SentenceTransformers all-MiniLM-L6-v2, 384-dimensional embeddings), an 8-section dynamic prompt builder that injects identity, emotion, RAG knowledge, RAG memory, and conversation history into each LLM call, and Server-Sent Events (SSE) streaming output that delivers identity and emotion metadata before the first LLM token. The language model is Mistral-7B-Instruct Q4_K_M GGUF, running locally via llama-cpp-python (n_ctx=2048, n_threads=8, max_tokens=256) with stop-token hallucination suppression and a 3-attempt retry strategy. Audio input is tran-scribed offline by OpenAI Whisper tiny. Text-to-speech output was deliberately omitted in the final build to eliminate 300–600 ms of synthesis latency and approximately 80–120 MB of peak RAM overhead on the 16 GB target machine. A psutil-based memory watchdog at 78% RAM usage triggers gc.collect() to prevent silent out-of-memory crashes
Zitieren
APA
Arivanantham, M. T., Salgar, H., Shaikh, B., & Waghchaure, S. S. A. A. (2026). PersonaX: Your Digital Reflection - A Locally Deployed, Biometric-Aware, Emotion-Conditioned Personalized AI Assistant. International Journal of Advanced Research in Science Communication and Technology, 362. https://doi.org/10.48175/ijarsct-36539
Vancouver
Arivanantham MT, Salgar H, Shaikh B, Waghchaure SSAA. PersonaX: Your Digital Reflection - A Locally Deployed, Biometric-Aware, Emotion-Conditioned Personalized AI Assistant. International Journal of Advanced Research in Science Communication and Technology. 2026:362. doi:10.48175/ijarsct-36539
BibTeX
@article{Arivanantham2026,
author = {Arivanantham, Mr. T. and Salgar, Hariom and Shaikh, Bushara and Waghchaure, Sohel Shaikh and Abhijit},
title = {PersonaX: Your Digital Reflection - A Locally Deployed, Biometric-Aware, Emotion-Conditioned Personalized AI Assistant},
journal = {International Journal of Advanced Research in Science Communication and Technology},
year = {2026},
pages = {362},
doi = {10.48175/ijarsct-36539}
}
EUVIMED