Technical Articles

Technical articles.

Practical notes from our work with private AI, local RAG, computer vision, and automation pipelines.

Diagram of orchestration failure modes and a four step method to simulate and test agent systems

Your multi-agent system is a distributed system

Evals score answers, not orchestration. How to find deadlocks, duplicate side effects, and partial failures with state machines, simulation, and Monte Carlo runs.

Read article

Diagram comparing a single agent with tools against agents split by permission scope

Split agents by blast radius, not by intelligence

Multi-agent architecture solves a coordination problem, and most projects do not have one. A practical rule for deciding when a second agent is actually justified.

Read article

Excel governance automation architecture for financial spreadsheet risk scoring

Excel governance automation for financial teams

How finance teams can keep Excel while adding read-only versioning, formula audits, risk scoring, and AI-assisted governance.

Read article

AI on-call agent architecture for Jira and AWS support automation

AI on-call agents for Jira and AWS support workflows

How small and medium companies can connect tickets, cloud signals, logs, runbooks, and past incidents to prepare repeated support investigations.

Read article

SELECT * FROM reality;

-- GenAI in Production

The Text-to-SQL reality: Engineering over Prompting

Text-to-SQL looks like magic in a 2-minute pitch. But what happens in production? We explore non-determinism, regional drift, and why enterprise AI requires strict LLMOps and air-gapped data governance.

Read article

Security camera view of a red car in a private garage with the license plate hidden

Offline AI security camera with computer vision, local RAG, and Gemma

An anonymized camera deployment configured for local video processing, private context retrieval, and a local model decision before alert routing.

Read article

Pipeline diagram showing local LLM sensitive data detection and deterministic masking

Using local LLMs to find sensitive data before masking

A safer pattern for realistic lower-environment data: local LLMs locate sensitive values, while deterministic systems replace them with consistent fake data.

Read article

Coming next

More deep dives we plan to publish.

Planned

On-Premise LLM Architecture

Model serving, GPU sizing, retrieval stores, access control, prompt logging, and monitoring patterns for open-weight private models.

Planned

AI Pipelines with Audit Trails

How to connect private model decisions to queues, databases, tickets, and human approvals while preserving traceability.

Need one of these systems before the article?

We can design the architecture, deploy the model stack, and connect the automation to your production workflows.

Get a Workflow Assessment