HCS-R-2026-04May 2026Video AI · Verticalisation
Content-Adaptive Verticalisation: Region-of-Interest Tracking for Automated 16:9 to 9:16 Reframing
Abstract — Converting widescreen video to vertical formats at scale requires deciding, frame by frame, what the shot is about. We present a content-adaptive reframing pipeline in which the tracked region of interest is conditioned on content class: speaker face saliency for interviews, ball-and-play tracking for sports, subject-agnostic animal tracking for wildlife footage, and automatic split-tile composition when two speakers share the frame. A temporal smoothing objective penalises crop-window jitter while preserving intent-critical motion. On a corpus of 12,400 broadcast segments across six genres, our method reduces manual reframing corrections by 78% relative to face-tracking-only baselines and sustains real-time throughput on live streams.
HCS-R-2026-02March 2026LLM Systems · Routing
Value-Per-Token: Preference-Aware LLM Brokering with Consolidated Benchmark Priors
Abstract — Model choice for a given prompt is usually made statically, per application, despite order-of-magnitude differences in cost and quality across the model landscape. We formalise value-per-token (VPT), a quality-adjusted cost metric, and describe a brokering architecture that classifies prompt intent, filters candidates by hard capability requirements (context length, modalities, tool use, language coverage), and ranks the remainder against a user preference vector spanning cost, accuracy, fairness, and latency. Public benchmarks are consolidated into per-task-type priors that are continuously recalibrated with online evaluation. In production traces spanning 12.4M routed prompts, preference-aware routing reduced spend by 58% while holding task-level quality within measurement noise of an always-frontier policy.
HCS-R-2026-01January 2026Narrative AI · Microdrama
Hooks and Cliffhangers: Narrative-Structure-Aware Segmentation of Long-Form Video into Episodic Micro-Series
Abstract — Repurposing long-form drama into short episodic series is dominated by manual editorial work: identifying the storyline, selecting gripping moments, and cutting episodes that respect scene continuity. We propose a narrative-structure-aware segmentation model that builds an intensity curve over the full runtime, detects and scores audience hooks (reveals, twists, confrontations), and solves for episode boundaries that open on a hook and end on a cliffhanger, subject to scene-boundary, dialogue-integrity, and music-bridging constraints. Human raters preferred our automated cuts over shot-boundary baselines in 83% of pairwise comparisons, and completion-rate telemetry from published micro-series shows parity with hand-cut editorial benchmarks.
HCS-R-2025-03November 2025Multilingual AI · Compliance
Native-Script Compliance Detection Across Ten Indian Languages: Beyond Translate-Then-Classify
Abstract — Compliance screening for broadcast content in India must operate across at least ten languages, four scripts, and pervasive code-mixing — conditions under which translate-then-classify pipelines systematically miss dialectal profanity, culturally specific slurs, and idiomatic threats. We train native-script detectors for fourteen compliance themes using a shared multilingual encoder with language-adaptive heads, and evaluate on a hand-labelled corpus spanning Hindi, Tamil, Telugu, Bengali, Marathi, Kannada, Malayalam, Gujarati, Odia, and Punjabi audio, text, and burned-in captions. Native-script detection improves recall by 21 points over translation baselines at matched precision, with the largest gains on code-mixed and dialect-heavy content.
