// the explorer
AI Explorer
By day, I am a Strategic AI Advisor and Enterprise Architect. I solve the challenge of integrating modern AI into rigid, legacy core systems in highly regulated industries. I design governed API layers, MCP-based agentic systems, and the cloud-native infrastructure required to make these models scale securely in production — across AWS, Azure, and GCP.
On the weekends, I run empirical experiments on LLMs and SLMs — adversarial tests, benchmarks, and compliance experiments on open-source models to see where they actually break. I replicate research paper findings on local hardware, focusing on structured output failures, adversarial guardrails, context position bias, and compliance enforcement. Real benchmarks. Real limitations. No hype.
I also build with agent frameworks — LangChain, CrewAI — and test how they hold up when wired to real cloud backends: Bedrock, Vertex AI, Azure AI Foundry. The goal is always the same: find what breaks before production does.
// Core Focus Areas
- ›Governed API & MCP Layers
Designing enterprise API ecosystems and MCP-based agentic systems that integrate AI safely into regulated, legacy environments.
- ›Agent Framework Testing
Evaluating LangChain and CrewAI agents against real cloud backends — AWS Bedrock, Azure AI Foundry, GCP Vertex AI — to surface production failure modes.
- ›RAG Compliance Enforcement
Designing multi-tier verification architectures that prevent LLMs from advice violation (e.g. money laundering or clinical bypass).
- ›Structured Output & Context Bias
Benchmarking JSON output modes under schema complexity limits and analyzing middle-context retrieval degradation in small open-source models.
- ›Prompt Injection Defense
Hardening models against system role prompt hijacking and jailbreaking using NeMo Guardrails and Llama Guard.
// Models in the Lab
*Focus is predominantly placed on offline-first, small parameter models run on local hardware to establish developer and cost constraints.
// Visual Testing Workflow
Applied Research
Translating academic paper methodologies and arXiv benchmarks into runnable local test code.
Agentic Engineering
Orchestrating complex test suites and pipelines using advanced coding agents (Claude Code, Gemini Assist).
Statistical Validation
Analyzing boundary conditions, response latency, and compliance rate across thousands of runs.
Open Disclosures
Publishing raw, honest failure modes and limitations rather than idealized happy-path demonstrations.