Teaching AIs to Reason and Code, Confidentially

Large Language Models (LLMs) have advanced rapidly, creating new opportunities for automating complex software-engineering tasks, yet today’s models still produce semantically flawed code and raise safety, privacy, and lock-in concerns on centralized clouds. I present an end-to-end framework that teaches AI to reason about code and executes it on a decentralized, privacy-preserving infrastructure. At the modeling layer, I orchestrate a quorum of specialized LLM agents. A Director LLM coordinates a concept agent rooted in programming-language theory, language-specific experts, and a compiler-driven feedback loop. Implementations such as UniTranslator and Smartify deliver state-of-the-art translation, synthesis, and vulnerability repair, especially for low-resource domains like smart contracts. At the systems layer, I introduce DeFaaS, a blockchain-managed, multi-cloud Function-as-aService platform that removes single points of failure. I further prototype OGAIS, which enables trusted, on-device LLM inference triggered and verified by smart contracts, and I demonstrate zero-knowledge-proof workflows that preserve user privacy. To sustain performance, I repurpose Tensor Processing Units as cryptographic accelerators, cutting the latency of homomorphic encryption and zero-knowledge proofs by an order of magnitude. The resulting stack keeps every model invocation auditable while sensitive data stay encrypted. Together, these contributions advance AI-driven software engineering and establish a secure path for its deployment. By uniting reasoning-centric agents with verifiable, decentralized execution, this dissertation lays the groundwork for autonomous development tools that are demonstrably more accurate, transparent, and trustworthy.

Name thesis:degree_name

Doctor of Philosophy

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Houston

Year dc:date.issued

2025

Author dc:creator

• Karanjai, Rabimba 1989-

Advisor dc:contributor.advisor

• Shi, Weidong

Committee members dc:contributor.committeemember

• Xu, Lei
• Wu, Panruo
• Huang, Shou-Hsuan Stephen

• Computer science

Language dc:language.iso

English

Handle dc:identifier.uri

https://hdl.handle.net/10657/20698

OAI identifier oai:identifier

oai:uh-ir.tdl.org:10657/20698