My Work: From Principled Research to Practical Implementation
Here is a selection of my major projects. Each is an exploration in building systems that are not just intelligent, but also transparent, auditable, and accountable by design. Click on any project to see a detailed case study of the research, architecture, and outcomes.
Governable Agentic AI
Architect
A self-aware, multi-agent system built for auditable growth, strict human primacy, and test-first execution.
Problem: Autonomous AI systems often act as opaque black boxes, executing actions and expanding scope without verifiable guardrails or deterministic rollbacks.
Approach: Built a self-growing agentic loop that isolates the LLM as a replaceable driver. Execution is sandboxed, file writes use atomic hashing for safe rollbacks, and new capabilities must pass a rigorous evaluation suite before promotion.
Outcome: A highly observable, self-reporting AI operator that proves autonomous systems can expand their own capabilities while remaining strictly governed by human (“Maker”) control.
Cognitive AI Analytics
NeuroTrace
A reproducible analytics framework that models the cognitive growth of AI assistants, turning conversation logs into research insights.
Problem: “Personalization” in AI is often a black box; we need structured methods to observe how an AI’s behavior actually evolves over time.
Approach: A deterministic CLI pipeline transforms raw conversation logs into analyzable interaction graphs, mapping messages to cognitive zones like Planning, Memory, and Feedback.
Outcome: Resulted in a formal research paper identifying key behavioral patterns, including a recurring
Feedback -> Planning corrective loop observed in ~26% of feedback events.Portfolio & AI Interface
MD Grid Engine
The component-driven content engine for this portfolio, designed to be the future UI for visualizing Architect's cognitive processes.
Problem: I needed a system to author component-rich content without a heavy CMS, and a visual layer to eventually render complex AI decision traces.
Approach: A custom Next.js engine with a runtime MDX compiler. A parser maps Markdown blocks to a registry of UI components like Grids and Cards.
Outcome: This portfolio website, which is fully powered by the engine and is ready to integrate with Architect as its primary user interface.
Agentic Writing & UX
Writing Agent
An agentic writing assistant that rewrites, condenses, and shifts tone through a small, production-minded agent loop instead of a one-shot prompt.
Problem: Most AI writing tools behave like opaque one-offs: they rewrite once, with no guarantees about clarity, faithfulness to the source, or tone-risky for high-stakes communication.
Approach: Built a compact agent that plans the goal, drafts an edit, scores it against a rubric (clarity, faithfulness, tone, length), and runs a single self-correction pass when scores miss acceptance thresholds, with all steps logged for later inspection.
Outcome: A demo-ready, production-aware writing assistant that delivers consistent, auditable rewrites and serves as a reference slice for Grammarly-style agents with evaluation and guardrails built in.
Cognitive Orchestrator Retrieval Architecture
CORA
A governed retrieval and orchestration system designed to turn complex knowledge work into traceable, evidence-backed outputs.
Problem: Most RAG systems retrieve text, but they do not reliably show why an answer should be trusted, how it was assembled, or how sensitive context is handled safely.
Approach: Designed CORA as a modular architecture with scoped retrieval, specialist routing, provenance tracking, verification layers, and observability-first system awareness so answers remain auditable and adaptable across local and cloud environments.
Outcome: Produced architectural design blueprints for a next-generation, evidence-first RAG system that prioritizes trust, traceability, and controlled intelligence orchestration.
Agentic Workflow & State Management
Communication Agent – Triage & Response System
A fully engineered MVP that connects to Gmail, uses an external LLM for triage and drafting, and manages ticket state (SQLAlchemy) via a comprehensive API and server-rendered UI.
Problem: Transforming raw, continuous email streams into auditable, stateful tickets requires a robust, production-grade agent system with secure authentication and state management.
Approach: Built a backend-first MVP using FastAPI and SQLAlchemy to manage ticket flow (NEW -> NEEDS_ACTION -> DONE), securely integrating Gmail via OAuth and abstracting LLM calls with a pluggable adapter.
Outcome: Delivered a production-ready, stateful system that automates the core support lifecycle, provides four real-time reporting boards (Jinja2), and adheres to strict security protocols for credential management.
Roadmap
Current Focus & Future Work
I’m currently building a connected portfolio of governed AI systems, observability-first tools, and evidence-backed agent workflows. The direction is moving from isolated prototypes toward a more unified stack where retrieval, governance, writing, analytics, and interface layers can work together as one coherent system.
Now Available:
- MD Grid Engine - the content and interface engine behind this portfolio
- NeuroTrace - the core cognitive analytics framework for tracing assistant behavior and interaction patterns
- Writing Agent - an auditable, self-correcting agent workflow for controlled text rewriting and tone transformation
In Progress:
- CORA - a governed, evidence-first retrieval and orchestration system focused on provenance, verification, and trustworthy knowledge workflows
Up Next:
- Connecting Architect, CORA, and the MD Grid Engine into live visual demos of governed AI workflows
- Extending observability and evaluation across projects so decisions, evidence, and failure modes stay visible by design
- Turning more blueprint-stage systems into runnable demos, release-ready repositories, and public technical documentation