Shakti: A Trauma-Informed Trilingual Women's Safety AI with Neuromorphic Affective Arbitration — Architecture, Dataset Design, and Training Methodology
Birla Institute of Technology and Science, Pilani
Abstract
We present Shakti-v1, the first deployment of a neuromorphic SNN affective core in a women's safety large language model. Shakti is built on Gemma 3 27B IT, fine-tuned via Vertex AI Managed Tuning (LoRA, adapter size 16, 3 epochs) on 2,000 human-authored trilingual seed samples spanning nine trauma-informed response categories. Four Leaky Integrate-and-Fire neurons — Bhaya (fear, τ=3), Vairagya (calm, τ=10), Shraddha (trust, τ=8), and Spanda (aliveness, τ=5) — arbitrate a four-tier safety response mode (SAFE, WATCH, ALERT, CRISIS) per message at inference time. A SHA-256 hash-chained ImmutableLog provides tamper-evident conversation evidence addressing India's BSA Section 65C gap. Deployed live at…
Citation impact
- FWCI
- —
- Percentile
- —
- References
- 21
Authors
1Topics & keywords
- Neuromorphic engineering
- Conversation
- Inference
- Software deployment
- Training (meteorology)
- Adapter (computing)
- Latency (audio)
- Gender equality