Jeremy Arancio

EPAM Systemsis an international technology and software engineering services company. As part of its collaboration with **SAP**, I was tasked with designing and deploying an enterprise AI solution for SAP's Communications team.

The objective: **implement a RAG system using internal news sources**.

I led the project as **Tech Lead**, from the POC phase to production deployment. The technical team consisted of:

1 Group Lead

3 Data Scientists

Key achievements:

🔷Production deploymentof an enterprise AI solution based on **Retrieval Augmented Generation (RAG)**. The product has nearly 10 users and continues to expand to other team members.

🔷 Development oflarge-scale data ingestion pipelines. 50+ documentsindexed daily via **Azure Databricks**. This infrastructure ensures instant access to up-to-date information.

🔷 **Accelerated development cycles**. Refactoring of the Back-End and automation of CI/CD pipelines led to a reduction in deployment time **from several weeks to a few days**, enabling rapid and continuous iteration.

Technical Context:

Python — Azure Databricks — CI/CD — Embedding Models — API — Clean Code — Docker — Git — GitHub — LLM — NLP — Machine Learning — Data Engineering

Alma is a Canadian startup aiming to make food tracking a breeze with AI. I joined the team as a Machine Learning Engineer to integrate and leverage open-source food databases, such as Open Food Facts or USDA.

The team consisted of:

In close collaboration with the CEO and CTO, I was able to:

🥗 Integrate open-source databases into the Postgres relational database. In total, over 2.5 million products were added to the startup's catalog. This also allowed the startup to become independent of paid APIs like FatSecret or Edamam.

🥗 Improve the RAG pipeline designed to return the correct food item based on the user's query. The new algorithm is more economical in terms of LLM calls and reduces response time.

🥗 Collaborate on Back-End development to meet the needs of iOS developers.

Open Food Facts is a French open-source organization. Its mission is to create the largest free and open food database in the world. To date, it includes over 3.5 million products in 150 countries. It is supported by over 12,000 active contributors worldwide.

I joined the core Open Food Facts team to develop and integrate ML functionalities as microservices. The technical team consisted of:

1 Full-stack engineer

1 ML engineer

1 DevOps

1 Mobile developer

Key achievements:

🍊 Ingredient spell checker: Errors in ingredients extracted by OCR compromised the analysis of product composition. I developed a spell-checking system by fine-tuning open-source LLMs. The solution is deployed on GCP and is 100 times cheaper than GPT-4o. This feature increased the number of identified ingredients by 11%.

🍊 Crop detection: I deployed an object detection model (YOLO-v8) on mobile using TensorFlow Lite. This feature improves user experience and speeds up the contribution process.

🍊 Database optimization: Using DuckDB and Arrow, I compressed the database from 43 GB to 6 GB. It is now easier for users and companies to leverage the open-source database. We use Hugging Face for hosting, improving accessibility and lightening server load.

🍊 Community engagement: I actively participated in initiatives to support community growth and collaboration.

Technical Context:

Python - Metaflow - DuckDB - AWS - Sagemaker - GCP - API - Clean code - Argilla - vLLM - Docker - Typer - Bash - GIt - GitHub - LLM - NLP - GPUs - Machine Learning - Data science