Deep dives into the architectures, decisions, and lessons learned from building AI systems. Each study covers the problem, solution, metrics, and what I would do differently.
Three interconnected platforms that form a complete enterprise AI ecosystem. AIthena provides the data foundation, Navari delivers agentic execution, and Agentis orchestrates them into a unified strategy.
FAIR data fabric powering the entire agentic ecosystem
Building the data infrastructure that makes enterprise AI actually work. Without clean, accessible, well-governed data, even the most sophisticated AI is useless. AIthena solved the foundation problem.
From a nights-and-weekends prototype to 1,000+ enterprise users
Building an agentic AI platform that delivers complete working solutions, not just answers. How MCP-based tool execution transformed how scientists interact with AI.
Unifying Data, Models, and Agents into enterprise AI strategy
The orchestration layer that answers: how do we make AI work at enterprise scale? Agentis connects AIthena's data fabric with Navari's agentic capabilities and domain-specific models into a coherent ecosystem.
One research kernel, adapted across three organizations, now running around the clock on borrowed iron
Most autonomous research agents are one-shot: loop, produce an artifact, exit. ARIA was built to run for months, accumulate signal, critique its own results, and recover from its own failures. The same kernel was adapted across three organizations (a global pharma R&D org, an enterprise setting, and SocialEyes retinal AI): 7 instances, 4 domains, 19,364 commits, 97.8% autonomous resumption, a filed provisional patent, and a written-up preprint. Its successor, ARIA-SE, now runs continuously on a borrowed NVIDIA 8xH100 node, 8,700+ sessions in.
Proving that focused training beats raw scale
Training oncology-specific multimodal models that outperform larger general-purpose models. Multi-teacher knowledge distillation at three scales: 4B, 1.7B, and 500M parameters.
9 agents debating over 8,000 securities daily
A fully autonomous trading platform where specialized AI agents with distinct investment philosophies debate positions through structured rounds. Running 24/7 on a Raspberry Pi.
Identity that travels with you across every Claude interface
Building a persistent identity layer for AI assistants. Continuum solves the amnesia problem—every Claude session starts with full context of who you are, what you are working on, and how you communicate.
128GB VRAM enabling experiments that rival funded research teams
Building a personal ML research infrastructure on NVIDIA DGX Spark. The GB10 Blackwell GPU with 128GB unified memory enables training and experiments that would otherwise require expensive cloud compute or institutional resources.
Extending Claude to social media, knowledge bases, and beyond
Building a collection of MCP servers that extend Claude capabilities to Instagram, LinkedIn, X/Twitter, Obsidian, and image generation. Publishing content becomes as simple as asking Claude.
How simple crypto bots evolved into a 9-agent autonomous trading system
The journey from single-strategy RL bots to multi-agent consensus trading. Three generations of autonomous trading systems, each building on lessons from the previous: Trading Agents → Chimera → Agora Quantum.
New case studies are added as projects mature. Want to see a specific deep dive?
Request a Case Study