Work

1. 

   Formula SpANow

   AI Engineer

   Jul 2026 – Present

   Working on an agentic Text-to-SQL assistant that plugs directly into the customer's ERP. A Next.js frontend is embedded in the ERP interface, inherits the user's ERP identity and calls a set of on-premise Python/FastAPI microservices. Users ask questions in natural language, and the agent generates, validates and executes SQL on the ERP database. It returns paginated, exportable results along with the generated query.

   • IAM service: authenticates users against the ERP user registry and issues JWT sessions.
   • GraphRAG retriever: business logics (question–SQL pairs) and annotated schemas in MongoDB, embedded in ChromaDB and projected into a Memgraph graph that encodes table relationships and table/column-level RBAC.
   • LangGraph agent: Devstral Small 2 24B (Q4_K_XL) served through vLLM, with tool calling, conversational memory and self-correction under deterministic guardrails and row-level security.
   • Knowledge-base manager: admin console and user-feedback loop that keep the business logics up to date.

   Key contributions

   • Designed the SQL validation flow: an AST-based guard (sqlglot) that enforces read-only, single-statement queries limited to the retrieved schema. Violations are fed back to the agent for self-correction.
   • Built the evaluation harness, with custom Text-to-SQL metrics across retrieval, SQL generation and database execution on gold datasets, integrated with Langfuse for end-to-end governance and traceability.
   • Introduced knowledge-base versioning and freezing on MinIO object storage (S3 API, boto3), which makes evaluation runs reproducible and comparable and keeps changes in the data out of performance comparisons.

   • LangGraph
   • vLLM
   • GraphRAG
   • FastAPI
   • Docker
   • MongoDB
   • ChromaDB
   • Memgraph
   • Langfuse
   • MinIO
   • sqlglot
   • Next.js
2. 

   EY

   AI Engineer

   Jun 2025 – Jul 2026

   • Built a RAG agent for IFRS 15 contract auditing using GPT-4o and ChromaDB, reducing analysis time by 70% and improving regulatory reference accuracy through automated compliance checklist generation.
   • Built a production enterprise RAG chatbot (FastAPI, LangChain, ChromaDB) for natural-language querying over 125 PPTX corporate documents; Azure Document Intelligence, slide-merge chunking, Cohere Rerank v4 Pro, and dual Azure/OpenAI GPT support.
   • Built a template-driven RAG system for automated field extraction from PDFs (native and scanned), using a 6-phase pipeline with hybrid retrieval, Cohere reranking and GPT-4o function calling to produce structured CSV output.

   • LangChain
   • FastAPI
   • ChromaDB
   • GPT-4o
   • Azure OpenAI
   • Azure Document Intelligence
   • Cohere Rerank
3. 

   EssilorLuxottica

   Data Analyst

   Sep 2023 – Jun 2025

   • Developed a Power BI dashboard for scrap monitoring, enabling targeted corrective actions that achieved a 2% scrap reduction and €500k in yearly savings.
   • Created a Power BI dashboard for quality control and productivity analysis, improving process supervision and reducing production-budget deviation by 3%.
   • Ran a SQL-based analysis of over 500,000 production records, identifying systematic inefficiencies and enabling a process improvement initiative that increased operational efficiency by 5%.

   • Power BI
   • SQL