From scratch
to scalable
digital systems.

Systematic research into reducing token consumption in large language model pipelines — through prompt restructuring, input compression, and context management strategies.

• ~72% reduction in token input achieved in controlled experiments
• Further optimization potential estimated at 85–92% under structured conditions
• Focus areas: prompt structuring, redundancy elimination, and dynamic context pruning
• Results are evidence-based from controlled test environments — not generalized claims

Research into securing AI system pipelines against adversarial input, prompt injection, and malicious manipulation — with a focus on production-grade AI agent architectures.

• Threat modeling for AI agent input pipelines
• Prompt injection mitigation strategies and pattern libraries
• Input sanitization and validation frameworks for LLM workflows
• Secure AI-agent architecture design with isolation and audit layers
• Malware-resistance patterns for AI-assisted code execution environments

An early-stage thesis exploring extreme visual data compression — investigating whether high-resolution images can be reduced to minimal storage sizes while preserving perceptual quality through intelligent reconstruction.

• Conceptual framework: 4K image → ~50KB storage → reconstructed into visually usable output
• Explores signal-based reconstruction rather than traditional lossy compression
• Early experimental stage — not validated for production use
• Research focus: identifying which visual information is perceptually essential vs. reconstructable