TL;DR - Emissary enabled Podqi to build a specialized classification language model that achieves 75% higher satisfaction than leading closed-source LLMs while cutting costs by 80% and eliminating their rate limitations— enabling Podqi to process millions of requests weekly to protect brands from IP infringement more effectively and affordably.

Introduction

Podqi is an AI-native platform that automates intellectual property protection at scale, scanning millions of potential infringements daily to help brands defend against counterfeits and unauthorized use. Backed by General Catalyst, Podqi protects major brands across entertainment, fashion, and consumer goods—recently removing 7,000+ counterfeit listings and blocking $4.5M in infringer revenue for clients like Hellstar.

Challenge

At the top of Podqi’s AI workflow is a critical component that determines whether a given infringement result has commercial implication - a decision that is made millions of times a day and must be accurate - since infringements with no commercial value are inactionable and would waste compute and generate noise if processed.

But generalized generative language models (LLMs) are not designed for such use-cases. Podqi found themselves constantly dealing with:

• Plateauing yet volatile performance as base models stop improving but remain hard to achieve consistency with.
• Rate Limits forcing them to juggle between models
• Exponentially increasing costs as they increased request volume and prompt size to handle failure modes.

Solution - Emissary Classification LM

To overcome these challenges, Podqi turned to Emissary to develop a specialized classification language model. Podqi's AI team fine-tuned Llama3.2-3b on the Emissary platform using only a few thousand samples collected from their dataset by leveraging Emissary's custom loss functions tailored for high-precision decision making.

Emissary handled all training infrastructure—GPU selection, access to the latest base models, templated training scripts, and model storage—allowing Podqi to focus squarely on their competitive advantage: their proprietary data and evaluation criteria. After testing 100+ configurations in under 4 weeks, they had their optimal model.

Podqi first deployed via Emissary's serverless inference for risk-free integration and experimentation at OpenAI-comparable pricing, then seamlessly transitioned to dedicated inference once they decided to scale and lower serving cost and latency. Throughout the engagement, Emissary's team provided hands-on support from architecture consultation to model optimization guidance, ensuring reliable performance at scale while maintaining Podqi's aggressive development timeline.

As a one-stop shop for training, deployment, and maintenance, Emissary enables continuous model retraining in just a few clicks, keeping Podqi's classifier performant as their data evolves. And if a better, new base model comes out, they can transition to that in hours - with built in knowledge cloning!

Impact

With their new, proprietary classification language model, Podqi achieved transformational results:

• Superior Accuracy: 75% win rate on disagreements when benchmarked against best-in-class closed-source models
• Unlimited Scale: No rate limiting—supporting millions of requests weekly with autoscaling infrastructure
• 80% Cost Reduction: 1/5th the price of their previous cloud LLM-based approach

Building on this success, Podqi is now expanding their use of Emissary's platform to other critical components of their AI pipeline in order to cement their place as the most technically sophisticated IP protection platform in the market.