AI Engineering & Automation

From Harness Engineering to the Two-Layer Interface Principle — operations run in conversation, interfaces step back to viewing decks

Multi-model orchestration & engineered AI systems

Autonomous systems that perceive, decide, and act

The craft of designing inputs for reliable outputs

Bringing AI into real workflows

The Two-Layer Interface Principle: conversation carries operations, interfaces stay read-only, documents govern the AI

Harness Engineering

Multi-model orchestration & engineered AI systems

Multi-Model Orchestration: route subtasks to the most suitable model based on each model's strengths.
Context Management: segment and summarize context during model switching to maintain coherence across long conversations.
Degradation & Retry: built-in fault tolerance in AI workflows so a single model failure won't interrupt the entire flow.
Performance Monitoring: track token consumption, latency, and output quality across models to optimize cost and speed.

Autonomous systems that perceive, decide, and act

Browser Automation Agent: migrated from Selenium to Playwright, demonstrating how Agents navigate complex web interfaces.
Custom Agent Pipeline: build Agents that combine LLM reasoning with API calls, file I/O, and structured data processing.
Tool Calling Pattern: design the Agent's ability to call external tools during decision-making.

The craft of designing inputs for reliable outputs

Structured Prompting: use system prompts, role definitions, and output format constraints to regulate model behavior.
Chain-of-Thought Reasoning: break complex tasks into step-by-step reasoning chains for improved accuracy and interpretability.
Few-Shot & Zero-Shot: choose appropriate examples or write precise instructions to get the best results without wasting tokens.
Iterative Optimization: systematically test prompt variants, measure output quality, and converge on the best expression.

Bringing AI into real workflows

Workflow Integration: connect AI pipelines to email, calendar, CRM, and other daily tools via API and Webhook.
Data Processing Pipeline: use AI to extract, transform, and analyze data from unstructured sources.
Human-in-the-Loop: design mechanisms where automation knows when human approval is needed.

The Two-Layer Interface Principle — software interaction architecture for the AI era

The conversation layer carries operations: CRUD, workflow triggers, and exception handling all run through natural language and Skills — every operation within reach of the AI.
The viewing layer stays read-only: graphical interfaces exist for people who do not operate — confirming state, reviewing results, demoing to outsiders. No write operations get built here.
The document layer governs the AI: rules, goals, ledgers, and logs live in documents as the single source of truth; the AI executes them literally, making documents the third interface — the one facing the AI.
Local-first: the source of truth is always a local file, interaction output flows into a backup-able pipeline, and clients hold copies and views only.

Engineering capability needs business context to create value