Homepage AI Video Assistant

AI Video Assistant

Headline Ask 200 hours of video a question. Get the clip., next to a laptop showing the Ask the archive chat and a card with the cited video clip

Services:

  • Product design
  • Full-stack development
  • AI engineering (RAG)
  • Video processing pipeline
  • Multi-tenant SaaS architecture

Industry:

  • Healthcare education · Professional training · SaaS

Technology:

  • Angular 21
  • NestJS / TypeORM
  • PostgreSQL 16 + pgvector
  • OpenAI GPT & embeddings
  • Google Gemini
  • PyannoteAI diarization
  • Vimeo API, ffmpeg
  • Tailwind CSS, Bun, Supabase

Challenge

  • “We have years of recorded lectures and panels. Nobody watches them, because nobody knows what is inside.”
  • “A confident wrong answer about a clinical topic is worse than no answer. We cannot have an AI that guesses.”
  • “We asked for a search box. What we actually need is a product: several libraries, several clients, our branding.”
  • “Who said what, and when? Our recordings have no transcripts, chapters or speaker names.”

Results

  • Every answer seeks the player to the moment it was said, at roughly four-second precision, so you watch 50 seconds instead of 56 minutes
  • 100% correct refusals on unanswerable questions in a 73-question benchmark and no hallucinated answers in evaluation, so “not in the archive” is an answer you can trust
  • 200+ recordings indexed with diarized transcripts, merged speaker profiles and chapters that read like a human wrote them
  • From a search iframe to a multi-tenant product: unlimited clients and libraries, tenant isolation enforced at the database layer
  • An automated evaluation suite scores every prompt and ranking change before release, so quality is measured, not assumed

Good fit for

  • Training providers, associations and universities sitting on large video libraries
  • Healthcare, legal and compliance teams that need answers with a source
  • Conference and event organizers who want recordings searchable and embeddable
  • Software companies that need a multi-tenant content platform without a rebuild

Someone asks you how accurate language models are at differential diagnosis. You know a speaker covered it. Was it the spring summit or the radiology panel? Forty minutes in, or ten? You used to open the recording and scrub. Now you type the question and get a two-paragraph answer and two clips, cued to 15:40 and 21:08, where the speakers actually said it. If nobody said it, the archive tells you so instead of making it up.

The question, the answer and the two cited clip cards from the real chat, above a session timeline with 15:40 and 21:08 marked

You ask. It answers. Then it shows you where. Every recording arrives with a transcript that knows who is speaking, a speaker profile that merges the same person across dozens of videos, and a chapter list that tells you in ten seconds whether an hour-long session deserves your time. Ask a question and the answer comes with numbered citations; click one and the player seeks to the moment the point was made, at roughly four-second precision. Type @ to ask one speaker, @@ to ask one session, or ask how a speaker’s view on a topic evolved across years of talks.

Dark infographic: question, relevance gate, retrieval, second gate, grounding contract, answer or honest refusal; with the figures 100% correct refusals and 0 hallucinated answers

It would rather say nothing than say something wrong. In a regulated field a confident wrong answer is worse than none, so the system is built to refuse. Two relevance gates stop off-topic questions before any model is called. Retrieval runs section-first, then passage by passage, so the answer comes from the right part of the right lecture. A strict grounding contract allows only facts present in the retrieved clips. On the 73-question benchmark the system declined every unanswerable question instead of guessing, and an automated evaluation suite with an LLM scorer re-checks every prompt and ranking change before it ships.

The real public portal in a browser card beside schematic wireframes of the client workspace and the agency admin, with the line Unlimited clients and libraries, tenant isolation at the database layer

A search widget was the brief. A product is what shipped. A branded public portal with embedded player, chapter navigation, filters, custom CSS and a domain allow-list for embedding. A client workspace for libraries, pipelines, speakers and settings. An agency-side admin for onboarding clients and watching the job queue. Underneath, a resumable seven-stage ingestion pipeline with two swappable transcription providers and per-library model settings, so quality is tuned without a redeploy. Unlimited clients and libraries, tenant isolation enforced at the database layer: a passive archive became an asset the organization can brand, embed and extend to new audiences.

If this sounds like your organization: you record everything and nobody watches it back. We make it answer questions, with receipts.

Let's get started!











    This site is protected by reCAPTCHA and the Google
    Privacy Policy and
    Terms of Service apply.

    Awesome!

    We’ll be right back for a good talk.
    Cheers ?

    Book a free consultation