Disclaimer: I created this piece of content for the purposes of entering the Gemini Live Agent Challenge hackathon.
The daily workflow in a food safety laboratory involves analyzing thousands of pesticide residue samples, often buried in complex Excel sheets or PDF reports. Finding specific analytical data usually takes time away from actual scientific work. I wanted to change that.
The Solution: LARS (Laboratory Analytics & Risk System)
LARS is a voice-first, multimodal AI agent designed to bridge the gap between analytical chemists and their complex datasets. Instead of writing queries, lab staff can simply use their voice or phone camera to ask about compliance limits or sample results, and get an instant, deterministic answer.
How It Was Built with Google AI & Google Cloud:
To ensure zero AI hallucinations—a critical requirement in public health and scientific environments—I designed a secure, two-tier architecture:
The Brain (Google Gemini 2.0 Flash Live API): I utilized the newest Gemini Live API to handle natural language understanding, real-time voice interaction, and multimodal vision capabilities (e.g., pointing the camera at a physical lab report). Gemini acts purely as the intent extractor and translator.
The Infrastructure (Google Cloud Run & Private VPC): The public-facing agent runs securely on Google Cloud Run. When a user asks a question, Gemini translates the intent and securely triggers a function call through a Private VPC to an isolated internal API.
The Deterministic Core (DuckDB): The internal API queries a structured, local DuckDB database containing real GC-MS/MS and LC-MS/MS analytical records. It returns the exact numbers, ensuring the AI only voices answers backed by strict data provenance.
By combining the conversational fluidity of Gemini with the scalable infrastructure of Google Cloud, LARS transforms a tedious manual task into a seamless, real-time conversation at the speed of thought.
Top comments (0)