AI / Media Infrastructure

An AI platform that turns decades of broadcast video into a searchable
knowledge base with transcripts, faces, chapters, and natural language search.

Python

FastAPI

Argo Workflows

Kubernetes

Qdrant

React

OpenAI

Anthropic

Overview

Project Overview

Built for a major broadcast archive, this platform auto-transforms raw video into searchable content via AI: transcripts, face recognition, chapters, summaries, and semantic indexing.

The Problem

Decades of footage were searchable only via manually assigned metadata, making content hard to discover. Each AI task—transcription, face recognition, chapter generation—ran independently, with undocumented, assumption-based dependencies. The platform orchestrates every AI service through one scalable workflow with real dependency tracing built in.

SOLUTION

Our Solution

Arithmiks designed the platform around a control plane and data plane architecture, keeping workflow coordination separate from media transfer for security and scale.

Tracing Reality Over Assumptions

Dependencies were assumed from service names, not real data flow. We traced actual reads/writes, rebuilt the pipeline, fixed bad dependencies, and unlocked blocked concurrency.

Visibility Into Long-Running Inference

AI stages ran minutes with no output, masking stalls as progress. Each video now runs isolated with live UI progress, so failures stay contained and retry without full restarts.

Search Built for How People Actually Ask

Raw video isn't searchable by nature. The final stage converts pipeline extracts into vector embeddings, powering NL search and a RAG chat interface over the archive.

HIGHLIGHTS