Transform 3-Hour Discourses into Profound Editorial Truth
Custom-built for Bhagavad Gita, Upanishadic dialogues, and Vedantic inquiry. Combines Domain-Primed ASR, 1D Temporal IoU Deduplication, and a strict Mathematical Meaning-Preservation Veto to eliminate sensationalist distortion.
"कर्मण्येवाधिकारस्ते मा फलेषु कदाचन — When action arises from clarity, not personal craving, fear vanishes."
Action Without Craving: The Essence of Nishkam Karma
Engineered for Deep Philosophical Accuracy
Unlike generic video clippers that optimize solely for shallow clickbait, this system was custom-engineered to preserve the nuance and completeness of Acharya Prashant discourses.
Meaning-Preservation Veto Engine
Enforces strict mathematical vetoes (S_preservation ≥ 0.68). Sensational soundbites that distort spiritual context are automatically rejected.
Hybrid Sanskrit & Semantic Search
Combines BM25 lexical token matching for exact Vedantic terms (कर्म, उपनिषद्, सांख्य) with 1024d pgvector cosine similarity via Reciprocal Rank Fusion (k=60).
Multi-Perspective Candidate Registry
Detects candidate windows across Q&A, Stories, and Philosophical Arguments, merging overlaps via 1D Temporal IoU (≥ 0.50) to prune 98% of compute.
PySceneDetect & Visual Cut Snapping
Snaps candidate start and end points to natural camera cuts, avoiding jarring mid-sentence or mid-shot visual glitches.
Single LLM Dispatcher & Qwen3-VL
Selective keyframe inspection and structured JSON analysis using local vLLM, eliminating high per-token cloud API costs.
Human-in-the-Loop Feedback Flywheel
Every editor decision (Accept, Reject, Amend) is stored in PostgreSQL to continuously train and fine-tune candidate ranking models.
The 6-Stage Intelligent Pipeline
From raw 4K video ingestion to mathematically verified, context-preserved editorial cuts.
Ingestion & Extraction
16kHz audio separation, 480p proxy creation, and keyframe sampling.
Vedantic ASR
Domain-primed Hindi/Sanskrit vocabulary transcription with timestamps.
Scene Transitions
Visual shot transition detection for natural cut-point snapping.
Semantic Chunking
Pause-aware semantic unit segmentation with 1024d Qwen3 dense vectors.
IoU Registry
Multi-perspective candidate deduplication pruning 98% of compute.
Veto Scoring
Multi-factor scoring with non-negotiable context preservation guardrails.
Ready to Process 3-Hour Discourses?
Extract profound spiritual insights, maintain non-negotiable Vedantic meaning preservation, and produce publication-ready shorts in minutes.