Shakti: A Trauma-Informed Trilingual Women's Safety AI with Neuromorphic Affective Arbitration — Architecture, Dataset Design, and Training Methodology

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 meetshakti.in. Training converged in 48 minutes (loss ~5.3 → ~0.5, zero truncated examples). Part 1 of a two-part series. Part 2 will report beta evaluation results. Maya-Defence Series Paper 3. ORCID: 0000-0002-3315-7907.