Abdoulaye S.

Architecture, development, and production deployment of an agentic platform that automatically generates analysis reports from a user-defined topic.

Business objective:

Drastically reduce human analysis time and document search costs while ensuring high reliability of results.

Key results:

• Very short analysis time thanks to the orchestration of 7 autonomous agents.

• LLM costs ÷4 (-72%) to date, via Redis semantic caching.

• 99.8% availability with a serverless Azure architecture.

• Infra cost ≈ $0/month (Container Apps vs AKS).

• Real-time experience via SSE streaming (no perceived latency).

Architecture & key decisions:

• Multi-agent orchestration (Planner / Researcher / Writer / Editor) with Draft → Reflection → Revision workflows.

• Multi-source tool calling (arXiv, Tavily, Wikipedia) with retry logic.

• Documented choices via 6 ADRs (semantic cache vs vector DB, SSE vs WebSockets, Container Apps vs AKS, agent temperature tuning).

• Automated deployment on Azure Container Apps via GitHub Actions CI/CD pipelines.

Technical Stack: FastAPI · GPT-4o · Upstash Redis · Azure Container Apps · Terraform · Docker · SSE · Application Insights · Python

Application:

Repository:

AI Assistant for a chain of 20 restaurants, used for customer support and daily operations.

Business objective:

Automate support at scale, drastically reduce customer errors, and eliminate superfluous infrastructure costs.

Key results:

• Error rate <2% (vs 15–20% baseline) thanks to a 4-layer anti-hallucination architecture.

• 5–10 ms latency via local FAISS engine → instant experience.

• Infrastructure cost = €0 (FAISS vs managed services).

• 99.5% availability in production (continuous monitoring via Render and UptimeRobot).

• Autonomous customer support: menus, hours, contacts, search by geographic proximity.

Architecture & key decisions:

• 100% agentic RAG architecture with 9 specialized tools, custom agentic orchestration (explicit sequencing of calls, fine-grained control of context and multi-step reasoning).

• Unique canonical knowledge base (KB) with multi-criteria consistency checks (price, hours, locations, dates).

• Geographic proximity search (Haversine).

• Automated scraping, full CI/CD, 54/54 unit tests passed.

• 6 documented ADRs (anti-hallucination strategies, agentic orchestration, LLM parameters, FAISS vs vector DB).

Technical Stack: FastAPI · GPT-4o-mini · FAISS · Docker · GitHub Actions · Pytest · Render · Python

Application:

Repository:

Mission carried out via Capgemini.

Design, development, and production deployment of a vector RAG system analyzing 30,000+ train journeys and technical publications to accelerate infrastructure engineering studies and reduce business operational load.

Business results:

• Analysis time reduced from several days to a few minutes via automated multi-pole report generation.

• API cost reduction through an embeddings/queries LRU cache (100 entries, 60 min TTL).

• Full automation: report generation and email sending via Azure OpenAI GPT-4o.

Architecture & key decisions:

• Hybrid RAG pipeline in production: BM25 + text-embedding-3-large embeddings + GPT-4o reranking (cross-encoder).

• Dynamic intent classification with Azure AI Search scoring profiles adapted to query context.

• Azure HA Architecture: PostgreSQL Flexible Server, Blob Storage, AI Search, Streamlit, secure network (VNet, Managed Identity).

• Technical leadership: code reviews, mentoring, definition of standards (Git workflow, DB nomenclature, SQL patterns).

Technical Stack: Azure AD · VNet · Managed Identity · Web App · PostgreSQL Flexible Server · Blob Storage · AI Search · LLM · Python