Ariel – Python, LLM, AWS

An AI-powered chat application designed to answer user queries by leveraging AgFlow’s structured and unstructured knowledge base. It provided agricultural professionals with contextual insights to support decision-making and future operations in the agricultural market.

Ariel was responsible for the following:

• created a PoC using Opensearch / Langchain / Streamlit;
• created backend architecture to work with web sockets;
• implemented RAG service from agricultural market reports;
• log all feedback in Dynamo DB;
• debugged and resolved the application issues;
• added metadata filtering;
• extended implementation to work with multi-agents for summarization / SQL query generator;
• integrated with AWS Bedrock services as Knowledge Bases / Prompt Manager / Agents;
• maintenance of the servers, with an availability of 99% of the time.

3rd project through an agribusiness market intelligence provider: a predictive service that forecasts next-day agricultural commodity prices, delivered via the AgFlow platform, API, or automated email for client risk management.

Among others, Ariel's contributions included:

• designed and built the full ML pipeline, from data extraction to client-facing inference;
• evaluated multiple imputation techniques and recurrent neural network approaches;
• maintained models through regular backtesting, retraining, and performance monitoring;
• containerized all pipeline components for scalability and reliability;
• deployed services on DigitalOcean (later migrated to AWS Fargate);
• exposed predictions via a REST API and automated client email delivery;
• ensured high system reliability with 99% server uptime.

A system that leveraged LLM agents and AWS services to automate the extraction and structuring of agricultural commodity prices from raw text and images.

Ariel's scope of contributions included, but was not limited to:

• designed a few-shot examples to guide LLM prompts for accurate data parsing;
• implemented a map-reduce approach to handle larger parsing contexts efficiently;
• extended the existing system to support LLM-based parsers alongside coded and manual parsers;
• integrated AWS Bedrock Agents and Prompt Manager to manage and store prompts across multiple environments.

2nd project through an agribusiness market intelligence provider: data migration project was a cloud migration initiative that involved transferring all existing data services and projects from Digital Ocean to AWS, ensuring high availability and preserving data integrity throughout the process.

Main responsibilities were:

• participated in planning the full migration timeline for data services;
• migrated Digital Ocean-managed PostgreSQL databases into custom containerized PostgreSQL instances to gain full control;
• configured AWS RDS Data Migration Service to synchronize Digital Ocean PostgreSQL with AWS RDS;
• set up AWS SNS to alert on synchronization failures and enable automated recovery;
• refactored and isolated structured GoLang data parsers for use in GoLang AWS Lambda functions;
• migrated unstructured GoLang data parsers to Python-based AWS Lambda functions;
• reimplemented GinGonic Golang REST endpoints as Python endpoints using AWS API Gateway;
• created and maintained materialized views to optimize data retrieval for web and mobile applications.

1st project through an agribusiness market intelligence provider: the project was a big-data geospatial pipeline that parsed satellite imagery and meteorological maps, aggregated climate indicators at the country level, and exposed the results via an API and interactive visualizations. The solution enabled programmatic access to processed weather data and delivered graph-based dashboards for analysis and decision-making.

Ariel's scope of work included the following contributions:

• built high-performance data extraction pipelines in GoLang;
• implemented OpenSearch for data storage and geospatial queries by regions;
• used Python (GeoPandas, Shapely) to aggregate satellite data and persist results in PostgreSQL;
• exposed aggregated datasets through PostgREST, later extended with GoLang and Python endpoints;
• maintained servers with a focus on 99%+ uptime and availability.