LIVESTOCKER Pro
LIVESTOCKER Pro is a product family supporting livestock production and operational workflows. I have worked on it as a full-time senior developer since 2018, with responsibilities spanning the core business application, mobile solutions, related websites, and the LIVESTOCKER Pro and Lite product variants.
Challenge
Livestock operations combine large volumes of interconnected business data with highly specialised workflows. Beyond daily operational tasks, the platform must support planning, optimisation, traceable reporting, and collaboration across several business areas.
The AI integration introduced an additional requirement: the assistant had to work with real, permission-controlled system data rather than provide generic answers. It also needed to recognise unavailable information and avoid inventing data that does not exist or cannot be accessed.
My role
The development team has previously included up to four developers and currently consists of two. The wider product team also includes a product owner, tester, support staff, and business stakeholders. As a senior developer, I carry broad end-to-end responsibility across web interfaces, applications, and product modules.
I mostly independently delivered the Hatchery, HR, variable-pay and settlement, Task Management, Dashboard, AI, and Egg Packing modules. I also developed numerous planning and optimisation interfaces within the Planning and Reporting areas, including workflows for fattening and breeding. My responsibilities also cover related UI and UX decisions, as well as web and mobile product development.
Solution
The delivered modules support daily operations, employee and task workflows, hatchery and egg-packing processes, and livestock planning and optimisation within one integrated product.
The AI assistant can work with 120 data and analytical tools and four knowledge bases. Its capabilities include finding partners and contracts, analysing livestock movements, mortality, FCR and daily weight gain, reviewing medication usage, and producing aggregations, averages, percentages, rankings, and comparisons.
The assistant cannot modify core business records. Persistent writes are limited to creating, updating, and deleting scheduled AI tasks, keeping its analytical capabilities within controlled operational boundaries.
Outcome and evidence
The product primarily supports customers in Hungary and additional European markets. A dedicated international website also supports the company’s distributor expansion. The screenshots in this portfolio use demo data.
In the documented tool-only TRL4 benchmark run on 11 August 2026, the active evaluation set contained 108 cases based on historical data, predefined questions, and expected answers. The LLM-as-a-Judge evaluation recorded 108 correct final answers and 108 correct tool chains, with no tool-execution errors and no fallback responses. One argument-validation discrepancy was recorded, but it did not affect the correctness of the final answer.
This result applies specifically to the documented 108-case tool-only evaluation set. It is not a general claim that every possible response produced by the assistant is error-free.
Technology
PHP 8.3 · SQL · REST API · Flutter · iOS · Android · OpenAI API · Gemini API · RAG · Tool calling · Codeception · PHPUnit · LLM-as-a-Judge evaluation
