AI Testimonial Intelligence Platform

Built a multimodal AI platform combining speech-to-text, sentiment analysis, and video understanding to automate testimonial insight extraction and story video composition at Cenphi.

Company: Cenphi  ·  Role: AI Engineer  ·  Year: 2022–2023

Problem

Marketing and sales teams collect large volumes of customer testimonial videos. Extracting compelling, usable insights — identifying the best moments, understanding sentiment arc, composing story narratives — was an entirely manual process requiring hours of human review per video. Cenphi needed an AI system to automate this pipeline end-to-end.

Technical Approach

Built a multimodal AI pipeline spanning speech, language, and video:

Speech & Transcription Layer

  • Implemented speech-to-text transcription pipeline using OpenAI Whisper with multi-speaker diarization
  • Speaker-attributed transcripts enable downstream analysis to track who said what and when

Language Understanding Layer

  • Sentiment analysis over transcript segments to produce a temporal sentiment arc across the video
  • Topic modeling and key phrase extraction to identify the thematic structure of the testimonial
  • Named entity recognition to extract products, features, and outcomes mentioned

Video Composition Layer

  • Designed the dynamic transcript strategy algorithm: given a target video length and composition goal (highlight reel, story arc, objection-handling clip), the algorithm selects segments by jointly optimizing over sentiment score, topic coverage, speaker diversity, and pacing
  • Automated generation of structured narrative scripts from selected segments
  • Video highlight extraction and assembly pipeline using FFmpeg

Stack

Python · OpenAI Whisper · spaCy · Transformers (HuggingFace) · FFmpeg · FastAPI · LLMs for narrative generation

Results

  • ~80% reduction in manual content review time per video
  • ~60% improvement in client decision-making speed for content selection
  • Shipped as the core AI feature of Cenphi’s commercial platform

Research relevance: This work sits at the intersection of my primary research interests — it required multimodal alignment (speech-language-video), temporal understanding, and evaluation of systems where “good” is inherently subjective. Designing the segment selection algorithm exposed me to the gap between engineering heuristics and principled multimodal optimization — exactly what video-language model research addresses.