Tessact · Case study
Keeping a seven-stage media pipeline moving.
At Tessact, I built and maintained a seven-stage asynchronous AI and media-processing pipeline. The work combined AI and computer-vision steps with media processing, with job tracking and recovery when steps failed.
- Role
- Software Engineer Intern
- Context
- February–May 2026 · Bengaluru
My contribution
I worked on the Python/FastAPI backend and asynchronous processing with Celery and Redis, as well as Docker-based computer-vision microservices on Google Cloud Run. My responsibilities included job-state handling, retries, idempotency checks, and long-running job recovery.
Making asynchronous work recoverable
A multi-stage pipeline needs to track more than whether a request succeeded. Jobs can run for a long time, a step can fail, and a retry can revisit work already attempted. I worked on state handling, retries, idempotency, and recovery so the pipeline could account for these situations.
Handling imperfect AI output
Some processing relied on LLM output that did not always arrive as valid JSON. I worked on fallback handling for malformed responses, alongside the surrounding processing pipeline. This was one of the boundaries where AI output needed to fit a structured backend workflow.
Tools in context
The work involved OpenCV and MediaPipe for computer vision, Gemini for AI processing, and FFmpeg and Remotion for media-related steps. FastAPI, Celery, Redis, Docker, and Cloud Run supported the service and execution layer.
What I delivered
During the internship, I built and maintained the pipeline’s job-state, retry, idempotency, and recovery handling. Company code and internal pipeline details are not public.
Technologies & references
- Python
- FastAPI
- Celery
- Redis
- Docker
- Google Cloud Run
- OpenCV
- MediaPipe
- Gemini
- FFmpeg
- Remotion