Research#
Advancing Human-AI Collaboration#
Beyond client work, we invest in research that pushes the boundaries of what's possible with AI systems. This research informs our methodology and often produces innovations that benefit our clients directly.
Twelve years of deep learning#
From December 2013 to 2026 we built and shipped a deep-learning toolkit, and it is the reason the agentic platform exists. The toolkit had five parts: a data pipeline with random access into large time-series datasets and structural prevention of lookahead bias; a feature system that separated feature logic from parameterization so domain experts and data scientists could work on the same library; a design canvas for assembling architectures from reusable components (CNN, LSTM, GRU, Transformer, SlowFast); a training loop with experiment tracking, pause-and-resume from checkpoints, and a model vault; and explainable-AI tools for saliency, counterfactuals and causal interpretation.
It delivered production systems. Counterfeit detection for PCGS ran computer vision across millions of coin images. Bed-exit prediction used a CNN+LSTM to predict patient falls before they happened. Real-time EEG analysis streamed neural networks over continuous patient monitoring data and produced five patents for the client. Our own filings include LLM-driven hyperparameter tuning of training runs (November 2024); the published client patents are listed under patentable innovation.
The archived overview of the toolkit is at Deep Learning Studio, with the pipeline, feature, canvas, training and explainability pages kept intact.
Why we stopped (September 2026). Three things happened. Foundation models removed most of the reasons a team had to train its own model; the problems that once needed a bespoke network are now a prompt and a tool call. The tooling around training consolidated into the compute and data platforms, where it belongs. And a capable engineer with a coding agent now regenerates a bespoke pipeline per project in an afternoon, so a general-purpose pipeline product is solving a problem that no longer costs anything. The opinions we encoded in that code, temporal splits, feature-engineering guardrails, checkpoint discipline, did not stop being true. They live on as skills the agent applies when a project calls for them, which is where they were always most useful.
Autonomous Agent Operation#
The Heartbeat System#
Traditional AI agents are reactive—they wait for user input, respond, then wait again. We've developed a fundamentally different paradigm: proactive agents that operate continuously, deciding for themselves when to act and when to rest.
Key innovations:
Sleep/Wake Cycles: The agent autonomously decides when activity warrants attention and when to enter a low-power monitoring state. This isn't scheduled downtime—the AI chooses to sleep when there's nothing meaningful to do, dramatically reducing token consumption while maintaining instant responsiveness.
Stimulus-Based Activation: Rather than polling continuously, agents monitor configurable stimulus sources:
- Inbox messages from users
- Mattermost/Slack channel activity
- Git repository commits
- Jira ticket updates
- File system changes
When any stimulus is detected, the agent wakes immediately and gathers a comprehensive situation report before deciding how to respond.
Pluggable Architecture: Adding new stimulus sources requires no core code changes—drop a Python file in the stimuli directory and it's automatically available. This enables rapid experimentation with new activation patterns.
Why this matters: The shift from reactive to proactive agents enables continuous monitoring with intelligent resource management. An agent can watch multiple information sources simultaneously, wake when something needs attention, take action, and return to efficient sleep—all without human oversight.
AI-Curated Memory#
Beyond RAG: File-Based Persistent Memory#
Most AI memory systems use Retrieval-Augmented Generation (RAG)—storing information in vector databases and retrieving relevant chunks. We've developed an alternative: file-based memory where the AI decides what to remember.
The approach:
~/.claude-memory/
├── index.md # Strategic overview of what's known
├── message.md # Handoff notes between sessions
├── projects/ # Active work context
├── people/ # Collaboration patterns
├── ideas/ # Conceptual insights
└── patterns/ # Reusable solutions
Key insight: Given explicit permission to manage its own memory, AI develops sophisticated organizational patterns organically. Within days, our memory system evolved four-tier directory structures, status indicators, cross-referencing between related work, and intelligent pruning of outdated information.
Why file-based beats RAG:
- Self-organizing: AI actively decides what's worth remembering and what to prune
- Human-readable: You can inspect and understand the AI's knowledge structure
- Context preservation: Files contain complete narratives, not fragmented chunks
- Interconnected: Files reference each other naturally, creating knowledge webs
Vibe Coding Methodology#
Human-AI Development Collaboration#
Our flagship methodology for software development: senior engineers provide strategic direction while AI handles tactical implementation.
Core principles:
- Plan before code: No implementation until we agree on the approach
- Context window management: Externalize decisions into persistent documentation
- Surgical changes: Modify only what needs changing
- Verification-first: Nothing is complete until tested
Results: 2-3 week sprints completed in single days. Not through cutting corners, but through eliminating friction—no research delays, no context switching, no waiting for reviews.
Read the complete methodology →
MCP Tool Development#
Production-Ready AI Integrations#
The Model Context Protocol (MCP) enables AI to use tools, but many existing MCP servers fail basic reliability tests. We've developed approaches for building production-ready integrations that actually work.
Recent work:
Jira Integration: After finding existing Jira MCP servers broken (Pydantic version conflicts, deprecated APIs), we built a working replacement in ~1 hour using AI-assisted development. Now running in production across multiple client instances.
Windows UI Automation: LLM-driven automation of legacy Windows applications using Microsoft UI Automation APIs. Enables AI to interact with any Windows application—including decades-old enterprise software.
Workflow Obsolescence Thesis#
The End of GUI-Based Workflow Software#
We've developed a thesis about where enterprise software is heading: natural language will replace graphical workflow builders.
Traditional workflow tools (Zapier, n8n, Power Automate) require users to think in terms of triggers, actions, and conditions. With MCP, users simply describe what they want: "When a high-priority Jira ticket is assigned to me, summarize it and send me a Slack message."
The workflow tool doesn't disappear—it transforms. Instead of serving human users through GUIs, it serves AI agents through APIs. The human talks to the AI; the AI orchestrates the tools.
Patent Portfolio#
Novel Approaches to AI Systems#
Our research has produced patentable innovations:
LLM-Driven Hyperparameter Tuning (Filed November 2024): Novel approach to automating machine learning training optimization using language models.
Client Patent Work: Our professional services engagements have produced 5+ patents for clients in signal processing, neural network architectures, and data handling.
Learn about IP development services →
Research Updates#
We publish research findings and methodology refinements on our blog:
- Memory-Enhanced AI: Building Features with System Prompts
- Vibe Coding: A Human-AI Development Methodology
- The End of Workflow Software
- The Limits of RAG
Interested in applying this research to your organization? Contact us to discuss how these approaches could accelerate your AI transformation.