P-01Multi-agent generation pipeline
Muqnea AI pipeline
The multi-agent pipeline behind a live generation platform: 82% lower cost from routing, caching and parallel steps, first output in 15–30s.
- Role
- Backend & AI Systems Engineer
- Company
- Muqnea
- Period
- May 2025 – Present
- Layers
- AI · Backend · Data
By the numbers
- lower generation cost
- 82%
- to first output
- 15–30s
- quality on the internal rubric
- 91%
- domain profiles, one pipeline
- 30+
- down from 2–3 days of expert work
- <10 min
01
Context
Muqnea is a live generation platform. It supports many sectors and company models — B2B, B2C and B2B2C — and all of them run through the same generation pipeline.
Generation is a chain of LLM steps, and the result goes to a product front end that has to render it. Free text is hard to render the same way every time.
Before this, one workflow took two to three days of manual expert effort.
02
What I built
- 01The backend and AI layer: Python microservices, queue-backed workers and the multi-agent pipeline that drives generation.
- 02Specialised agents that hand off schema-validated structured output, so everything downstream renders the same way every time instead of being parsed back out of free text.
- 03Routing that sends each step to the smallest model that passes its eval, caching for stable prompt prefixes, and steps that run in parallel.
- 0430+ domain profiles, so one pipeline serves every sector and company model the platform supports, with no per-client engineering.
- 05The platform API: versioned REST endpoints with authentication and rate limiting, used by the product front end and by internal services.
- 06An architecture for scale: stateless services, queue-based workers, Redis caching and partitioned PostgreSQL.
03
Results
- Generation cost down 82%, from three changes together: model routing, prompt caching and parallel steps.
- Time to first output down to 15–30 seconds.
- Quality measured at 91% on the internal rubric.
- A workflow that took two to three days of manual expert effort now finishes in under ten minutes.
- High daily generation volume, on a platform sized for a user base in the millions.
04
Stack
- Python
- Microservices
- Multi-agent LLM
- Structured output
- Model routing
- Prompt caching
- Queues & workers
- REST API
- Redis
- PostgreSQL
Details that belong to clients or employers stay private. Everything here is from my CV.
More work
All workP-03August 2022 – March 2024
Shared core of an enterprise ERP platform
Minaret ERP
AL-MOTKAMEL Global Co.
Case studyP-05May 2024 – May 2025
Records, scheduling, billing and reports
Clinic management system
Freelance · healthcare client
Case study