MCQ Scan

Context and problem

Advertising agencies needed an agentic ad/ops automation system. My work connected no-code workflows, LLM orchestration, product interfaces and deployable infrastructure.

Ownership and contribution

AI Research Engineer / Lead AI Engineer, 2024–present. Led architecture, implementation and deployment, coordinating a two-person engineering team.

Responsibilities

• Led development from API architecture and frontend integration through deployment.
• Built CI/CD deployment paths with Cloud Run, Secret Manager, Artifact Registry, Docker and GitHub Actions.
• Evaluated and integrated constrained-decoding tools, selecting LM Format Enforcer for enum enforcement; added Pydantic business-coherence validation.
• Optimized end-to-end workflows using dependency-aware parallelism, smaller-model routing and caching; fine-tuned Llama 3.1 8B with LoRA for Store Chat.
• For a client-infrastructure deployment, compiled the code into runnable software so delivery did not require handing over source code.

Architecture and data flow

The platform combines no-code client workflows, API and frontend integration, multi-agent orchestration and structured-output LLM pipelines. I architected and deployed the platform on GCP using Cloud Run, Docker, GitHub Actions, Artifact Registry and Secret Manager. This is a high-level account; client topology and implementation are private.

1. No-code client workflows
2. API and frontend integration
3. Multi-agent orchestration
4. Structured-output LLM pipelines
5. GCP deployment

Engineering decisions

Reliability, security and evaluation

• Constrained decoding and structured-output pipelines enforce schema-valid interfaces. Schema validity does not establish semantic correctness.
• CI/CD and GCP Secret Manager support deployment and secret handling.
• Client source, screenshots, evaluation datasets and deployment manifests are not included in this public account.

Outcome and status

Professional work documented in my CV evidence database. The reported latency improvement has no public benchmark dataset or baseline methodology attached. This case study summarizes my contribution; it is not a public demo or independent production audit.

Technology groups

Store Chat

A clothing-shopping assistant with Llama 3.1 8B LoRA, local vLLM inference, and Pydantic business validation. I evaluated constrained-decoding tools and integrated LM Format Enforcer to enforce enum values. Internal evaluation separates 97% structurally valid, enum-compliant requests from 95% semantic accuracy per attribute.

These are internal evaluation results, not a public benchmark. My contribution was tool evaluation and integration, not a custom sampler.

AgentDiet

I conceived an agent-governance project at MCQ Scan: signed Biscuit tokens, attenuated delegations, and local Datalog checks for tool permissions. I built a Python/FastAPI demonstrator and prepared the technical proposal, market analysis, budget, and project plan. AgentDiet received a 2026 Hyper Open X grant funded by France 2030.

The demonstrator validates simulated authorizations; MCP/A2A integration is the project target.

Latency and delivery

Reduced end-to-end workflow processing time by 60% through a parallelizable dependency graph, routing simpler steps to smaller models, and caching repeated calls (internal measurements).

Also built custom request tracing for per-step tokens, latency, cost and tool calls, plus availability alerts. Image-generation work covered Qwen-Image-Edit-2509 and ControlNet, GPU deployment on Vast.ai, and reverse-proxy access.

Proof and source material