NDA Guardian
Team led by a Google AI software engineer and IIT Bombay CS grad specializing in Gemini integrations, C++, and shipping large-scale productivity tools.
YouTube Video
Project Description
NDA Guardian analyzes employment contracts and NDAs locally on-device using FunctionGemma via Cactus Compute, escalating to Gemini only when external legal reasoning is required — ensuring sensitive documents never leave the device.
The core is an intelligent routing agent powered by FunctionGemma’s tool-calling. Every query is classified by a route_query() tool that evaluates query type, local confidence score, and clause sensitivity to decide on-device vs. cloud — with every decision logged and auditable. Extractive queries (“Who are the parties?”, “What is the non-compete duration?”) run entirely on-device via Cactus Compute in ~400ms. When escalation is necessary, only a 40–80 word clause summary is sent to Gemini — never the raw document.
The agentic pipeline chains five tools: route_query() → extract_local() or summarize_clause() → optionally escalate_gemini() → answer_with_citation(). FunctionGemma orchestrates this chain dynamically on-device, making routing decisions query by query.
Technologies: FunctionGemma via Cactus Compute SDK (Apple Silicon MPS), Gemini 1.5 Flash (Google AI SDK), Python FastAPI
Metrics: Tool-call correctness is auditable in real time via logged routing decisions. Speed is demonstrated by ~400ms local responses vs. justified cloud calls. Edge/cloud ratio targets 70%+ local. Privacy is anchored by a live counter showing zero raw document bytes sent to cloud.
Prior Work
100% during hackathon.