Safety Manager Assistant
A production AI co-pilot for fleet safety managers — product documentation, drivers data, alerts, coaching, device health and actions in one conversation.
Project Snapshot
Highlights
Named inventor on a filed patent (PCT WO 2025/064877 A1) for the multi-step, two-model database-querying method that powers the co-pilot's natural-language access to driver data.
Owned design and delivery of SAM — Driveri's Safety Manager Assistant — a production generative-AI co-pilot now used by thousands of fleet safety managers.
Built the natural-language query engine: a first model pulls a scoped subset from the main driver database into a temporary table, a second reads it — so 'which drivers had the most distracted-driving events this month?' returns in seconds.
Grounded product-documentation answers in RAG over Elasticsearch, kept current by an automated documentation-ingestion pipeline.
Stood up a human-in-the-loop labeling tool that turns reviewer judgments into a regression set, improving answer quality over time.
About the Company
Netradyne builds Driveri, a vision-based AI dashcam and video-telematics platform that analyzes nearly the entire driving day to help commercial fleets coach drivers and prevent collisions — used across thousands of fleets and hundreds of thousands of drivers. The Safety Manager Assistant (SAM), a generative-AI co-pilot shipped inside Driveri, sits on top of that platform: one conversational surface over product documentation, driver data, coaching, alerts, and device health.
Scope of Work
Design & delivery
Owned the co-pilot end to end — from the conversational surface for safety managers to a streaming FastAPI service rolled out to production.
Natural-language query engine
Designed the multi-step, two-model querying method — main-database retrieval into a temporary table, then a refined read — later filed as a patent.
Retrieval & documentation Q&A
Built the RAG layer over Elasticsearch that grounds product-documentation answers in the current docs.
Documentation-update pipeline
Automated ingestion that re-chunks and re-embeds changed documentation so retrieval stays fresh without manual re-indexing.
Labeling & evaluation flywheel
Stood up the human-in-the-loop labeling tool and turned reviewer judgments into a regression suite driving continuous quality gains.
The Challenges
Product docs, driver data, alerts and coaching lived in separate systems
Unified them behind one conversational co-pilot, so a manager asks once instead of stitching multiple tools together.
Answering questions over large driver datasets in plain language
A first model pulls a scoped subset from the main driver database into a temporary table, and a second reads that bounded set — the method later filed as a patent, avoiding one blind text-to-SQL call over a huge driver dataset.
Keeping documentation answers current as the product changed
RAG over Elasticsearch grounds answers in the docs, and an automated ingestion pipeline re-chunks and re-embeds updated content so the co-pilot reflects the latest version.
Measuring and improving answer quality over time
A human-in-the-loop labeling tool turns reviewer corrections into a growing evaluation set, so prompt and retrieval changes are checked for regressions before shipping.
Outcomes
The multi-step database-querying method was filed as a patent application, published as PCT WO 2025/064877 A1 (applicant Netradyne), with Adarsh Saini a named inventor.
Shipped to production and adopted by thousands of fleet safety managers.
Runs as a FastAPI service with LangChain + OpenAI orchestration, multi-step database-querying and retrieval over Elasticsearch.
An automated documentation-ingestion pipeline and human-in-the-loop labeling loop keep answers current and continuously improving.
Tech Stack
Have a project like this?
Tell us what you're trying to ship. We'll tell you honestly if we're the right studio for it.