The Ministry of Economy, Ecology and Agriculture of Ukraine digitizes licensing process with Gemma
Navatech
Bringing critical safety knowledge to frontline workers with EmbeddingGemma and Gemma 3n
Navatech delivers safety and operational knowledge to disconnected worksites through a mobile-first RAG platform
Navatech is an SaaS company building tools for high-risk industries such as construction, facilities management, manufacturing, and mining. A member of the Google for Startups AI First Cohort, they aim to provide frontline workers with immediate access to critical information, often in environments where connectivity is limited, expensive, or nonexistent.
To address these challenges, Navatech developed offline document intelligence that can work fully offline, integrating EmbeddingGemma for retrieval of relevant information and Gemma 3n to generate context-aware responses.
Enhancing field safety through offline document intelligence
For active worksites, a primary challenge is the tendency to communicate using casual terminology, whereas regulatory compliance documents are written in dense, formal prose. This friction is amplified when workers speak a different language than the original source text. Navatech uses EmbeddingGemma to create a “semantic bridge” that aligns real-world site conditions with safety protocols.
For example, a worker on a construction site may ask a safety-critical question in Arabic.
In a keyword-based search, the worker would not find any responses from the English-language documents which classify sparks as “Ignition Sources” and paint cans as “Combustible Materials.”
Because embeddings map conceptual relationships rather than string characters, they can be used to recognize the semantic intent of the worker’s query and retrieve the relevant safety information. By running EmbeddingGemma locally, a search through hundreds of documents can be orchestrated right on a worker’s phone, all in a matter of seconds.
Deploying RAG systems at the edge
Navatech’s solution to improve worksite safety is split into two related pipelines, separating heavy document processing from real-time, on-device search and synthesis.
First, company materials are parsed into clean, structured JSON then preprocessed and tokenized using the tiktoken library into overlapping chunks of 256 tokens. To ensure retrieval-friendly contexts, each chunk is processed in two ways:
- Semantic encoding: Chunks are vectorized into dense embeddings using EmbeddingGemma with SentencePiece tokenization.
- LLM-generated description: A reasoning model synthesizes concise descriptions and tags for each chunk to handle terminology gaps.
The compiled inputs—including IDs, raw text, vector embeddings, and synthetic metadata—are assembled into a highly compressed sqlite-vec database file and compiled directly into an Android application package (APK).
Once loaded onto a worker’s device, the system can operate completely disconnected from the internet. EmbeddingGemma generates a vector when a question is asked, and the application performs a similarity search over the local vector database.
Top-k rankings are calculated using a weighted sum of two scoring streams: a Similarity Score to evaluate the spatial proximity of the query vector relative to document vectors, and a Keyword Score to run BM25 token-matching over the plain text index. The highest-scoring document chunks are extracted and assembled alongside the worker’s original prompt. Gemma 3n then follows specific system prompts and synthesizes a grounded safety answer in the user’s language.
This approach prioritizes relevance and responsiveness over open-ended generation, allowing the system to handle fragmented data and varied terminology across different teams. By shifting these capabilities to the edge, Navatech reduces dependence on continuous connectivity and minimizes bandwidth requirements, making knowledge access significantly more resilient for low-connectivity environments.
Navatech exists to make AI work for the people who keep the world moving. Our mission is to give frontline workers practical, always-available support at the point of need—especially in environments where digital access is uneven, systems are complex and connectivity is limited.
Delivering back-office knowledge to worksites
Navatech measures success through operational signals like retrieval relevance, response speed, and feedback on whether the retrieved content ultimately helps teams complete tasks more safely and efficiently.
Looking forward, the team plans to deepen its edge-first AI stack with a focus on improving multimodal interactions across voice, image, and video to ensure the platform remains useful across all communication mediums. “We’ve built AI for the realities of frontline work — where cost, connectivity, and usability matter just as much as technical capability. Our focus is on delivering semantic understanding in a way that is practical, reliable, and accessible in the environments where it is needed most,” says Mukund Hirani, co-founder & CTO of Navatech.
This shift away from desk-based workflows ensures that critical safety knowledge is always available, regardless of how remote a worksite may be.
More from the Gemmaverse
Adaptive ML trains Gemma 3 for exceptional multilingual results
Quarks improves user experiences with Gemma 2 and Gemma 3
Sarvam AI built a translation model with Gemma 3 to translate all 22 officially recognized Indian languages
Institute of Science Tokyo creates powerful Japanese-focused LLM with Gemma 2
Introducing GAIA, a Brazilian Portuguese Gemma 3 model developed with ABRIA, CEIA, Nama, and Amadeus AI
