Insights
How we think, in writing.
Technical notes on building GenAI systems that survive production. Written for the engineers who build them and the leaders who commission them.
Architecture
RAG architecture for enterprise: what actually works in production
14 min
Most RAG implementations look fine in demos and fall apart in week two. The architecture decision that changes that.
Architecture
Document intelligence: the architecture behind processing 10,000 contracts a month
15 min
PDFs, scanned invoices, legacy formats — the full stack for extracting structured data from whatever documents arrive.
Architecture
Fine-tuning vs RAG vs prompting: a decision tree for enterprise teams
15 min
The answer is almost always prompting first, RAG second, fine-tuning rarely — and the specific cases that break the rule.
Architecture
Vector databases for enterprise: Pinecone vs Weaviate vs pgvector
13 min
What production deployments actually reveal about each option, beyond the comparison-post consensus.
LLM stack
GPT-5.5 vs Claude vs Gemini 3.1: a practical decision framework
12 min
Benchmarks are useless for your use case. How to actually choose models for business automation.
LLM stack
Prompt engineering for business workflows (not chatbots)
12 min
Prompts for internal automation are a different discipline from prompts for chat — higher stakes, subtler failure modes.
Agents
Agent orchestration patterns: when single-agent fails and multi-agent wins
16 min
Single-agent is underrated and multi-agent is over-hyped. A decision framework from production patterns.
Agents
The human-in-the-loop pattern: where AI needs a checkpoint
11 min
Full automation isn't always the goal. A practical framework for deciding where humans stay in the loop.
Deployment
The real cost of running LLMs in production
11 min
Token costs compound fast at scale. Four changes that keep deployments from blowing the budget.
Security
LLM security in regulated industries
14 min
Banking, healthcare, insurance — the compliance requirements are real, and standard deployment advice skips most of them.
These patterns are what we build.
If one of them maps to a process your team runs by hand, bring it to a working session.
Book a working session