Policybazaar constructed an advanced, multi-model parallelized orchestration pipeline. Media assets are now hosted securely within Cloud Storage before a Python-based orchestrator splits long video files into logical segments for processing. The foundational breakthrough of this architecture relies on a custom sequential accumulation logic. Rather than deploying rigid algorithms that search for an elusive, flawless ‘hero frame,’ Policybazaar instructed the Gemini models to continuously track and accumulate visual evidence of edges (Top, Bottom, Left, Right) over the entire duration of a camera pan to ensure the full windshield is visible.

This architectural shift allowed the company to effortlessly track over 30 separate parameters while discarding blurry or unreadable segments just like a human inspector. To balance infrastructure costs against performance constraints, the team deployed a highly efficient multi-model strategy. Simpler tasks like document parsing or basic object extraction run concurrently through models like Gemini 2.5 Flash and Gemini 2.5 Flash-Lite. Complex visual parameters are routed directly to Gemini 3 Flash Preview, while Gemini 3.1 Pro provides deep meta-analysis.

The team also utilized direct content injection, embedding specific inspection instructions directly alongside the video inside the User Prompt rather than System Prompts to secure tighter visual-logic grounding. The engineering team implemented Embedded Reasoning through strict Chain-of-Thought prompts, and the model is forced to define contextual boundaries and write out a detailed description of the camera path inside a designated JSON field before it is permitted to deliver a final true-or-false validation status. This extra layer of accountability significantly improved system accuracy, while optimizing hyperparameters to a temperature of T=1.0 enhanced the model’s grasp of fluid visual motion.

Policybazaar saw a dramatic accuracy shift from a 0-60% baseline to 100% on tested datasets by deploying Gemini models on Google Cloud. The multi-model pipeline dropped average inference latency to 11–13 seconds per video evaluation, crashing total advisor review loops from 15 minutes to just 1 to 3 minutes. These metrics directly optimized the consumer experience, pushing the five-minute policy decisioning window from 71% to 96%. For the Pay As You Drive (PAYD) segment, turnaround latency plummeted from 97% to 0%, securing instantaneous issuance for last-minute shoppers.

The removal of the administrative backlog successfully reclaimed 50% of advisor capacity, with clear milestones to hit 50% in the near term. Freed from the manual work of inspecting windshield videos, advisors can scale their output up to three times the current volume and pivot toward strategic tasks.

Crucially, this operational shift preserves the human element in insurance. Because insurance touchpoints frequently occur during high-stress life events like lapsed coverage or post-accident vehicle damage, advisors can now leverage automated root-cause case histories to bypass paperwork and focus their conversations on human empathy and rapid crisis resolution. In the upcoming phase, Policybazaar will continue co-architecting its ecosystem with Google Cloud, deploying real-time edge guidance models that actively correct customer recording errors at the source to secure an enduring competitive advantage.