Poster

Trainwise-AI

Inspiration

Educational and training content is increasingly spread across long videos, presentations, PDFs, and documents. Finding one specific explanation often means scrubbing through an entire video or manually searching across multiple resources.

We wanted to make training and learning content searchable, conversational, and instantly accessible.

Trainwise-AI turns videos and learning materials into a unified AI knowledge base where users can ask natural-language questions, receive grounded answers, and jump directly to the relevant source or moment in a video.

What it does

Trainwise-AI is a Retrieval-Augmented Generation (RAG) platform for video and learning content.

Users can organize content into lessons and add:

YouTube videos
PDFs
PowerPoint presentations
DOCX documents
Markdown files
TXT files

For videos, Trainwise-AI automatically downloads the content, transcribes the audio using Whisper, generates a summary, chunks the transcript, creates embeddings, and indexes the information in FAISS.

When users ask a question, Trainwise-AI searches the selected lesson's knowledge base, retrieves the most relevant information, and provides it to a locally running LLM.

Instead of only generating an answer, Trainwise-AI also provides the relevant source and video timestamp, allowing users to verify the answer and immediately return to the original material.

How we built it

Trainwise-AI uses a multi-stage local AI pipeline.

Content Ingestion

YouTube videos are downloaded using yt-dlp, while documents, PDFs, and presentations can be uploaded directly into individual lessons.

Speech-to-Text

Video audio is transcribed locally using OpenAI Whisper. Timing information is preserved so retrieved transcript segments can be connected back to their original location in the video.

Summarization

We use Hugging Face Transformers with facebook/bart-large-cnn to generate concise summaries of uploaded videos.

These summaries provide broader context when answering questions about the content.

Embeddings & Vector Search

Transcript segments and uploaded learning materials are converted into semantic embeddings using all-MiniLM-L6-v2.

The embeddings are stored in FAISS, allowing fast similarity search across the lesson's knowledge base.

Retrieval-Augmented Generation

When a user asks a question, the pipeline becomes:

Question → Embedding → FAISS Search → Relevant Context → Local LLM → Grounded Answer

The retrieved context and content summaries are passed to an Ollama-hosted LLM, which generates an answer grounded in the uploaded material.

Timestamped Video Retrieval

When information comes from a video, Trainwise-AI generates a timestamped YouTube link such as:

youtube.com/watch?v=...&t=125

Users can immediately jump to the exact part of the video where the information was discussed.

Tech Stack

Python
Flask
FAISS
Sentence Transformers
OpenAI Whisper
Hugging Face Transformers
facebook/bart-large-cnn
Ollama / Llama 3.2
yt-dlp
FFmpeg
React
ngrok

Challenges we ran into

One of our biggest challenges was connecting retrieved transcript chunks back to meaningful moments in the original video.

Traditional RAG systems retrieve text, but Trainwise-AI needed to preserve timestamp information throughout transcription, chunking, embedding, retrieval, and answer generation.

Another challenge was supporting multiple types of learning resources while keeping individual lessons isolated. We designed each lesson as its own FAISS-backed knowledge base so users can search a specific topic without retrieving unrelated information.

Running the AI pipeline locally also required balancing model quality, inference speed, and hardware requirements.

Accomplishments that we're proud of

We're especially proud that Trainwise-AI goes beyond simply summarizing videos.

It creates a searchable AI knowledge layer over entire lessons and training resources.

A lesson can contain videos, PDFs, presentations, and documents, and users can search across all of them through one natural-language interface.

We also built the core AI pipeline around locally running models:

Whisper handles transcription
Sentence Transformers generate embeddings
FAISS handles semantic retrieval
BART handles summarization
Ollama handles answer generation

This architecture provides greater privacy while reducing dependence on external AI APIs.

What we learned

Building Trainwise-AI helped us understand how the individual components of a complete RAG pipeline interact.

We worked with:

Semantic retrieval
Vector databases
Embedding models
Speech-to-text
Local LLM inference
Context construction
Video timestamp alignment
Multi-format document ingestion

One of our biggest takeaways was that an effective RAG system depends just as much on retrieval quality and data organization as it does on the LLM generating the final response.