Pseudo: Simulation of Omni Model Behavior

Chhetri, Alish

Pseudo: Simulation of Omni Model Behavior

Persistent URL

https://hdl.handle.net/10456/59221

Author(s)

Chhetri, Alish

Date Issued

May 6, 2025

Abstract

The artificial intelligence landscape today presents two divergent approaches: monolithic “omni-models” pursuing general capabilities through massive scaling, and specialized models that excel within specific domains. This thesis introduces Pseudo, a system that explores bridging these approaches through specialized model routing. Pseudo investigates how specialized models might be orchestrated while maintaining a unified interface. While the theoretical framework proposes sophisticated content analysis for routing requests, the current implementation uses deepseek-r1 through Ollama to classify user inputs and determine appropriate modalities. During development, the fragmentation of AI service provider APIs emerged as a significant challenge, leading to the creation of APICenter, a universal abstraction layer addressing the broader “provider abstraction problem” faced by developers working with multiple AI services. This research examines both the theoretical potential of intelligent routing to specialized models and the practical challenges encountered in early implementation. The findings suggest that while the current implementation has significant limitations, the approach of intelligent orchestration of specialized components could potentially offer advantages in resource efficiency, provider flexibility, specialized task performance, and ethical diversity. This represents an alternative direction to the prevailing assumption that increasingly large models are the optimal path toward advanced AI capabilities.

Major

Software Engineering

First Reader(s)

Graber, Emily