Insights

Research notes for the operating layer of production AI.

Sharp essays, decision notes, and field observations on infrastructure, agents, models, evaluation, governance, GTM, and the systems that make AI useful in real organizations.

Featured note

Why AI architecture is now an operating decision.

Infrastructure, agents, models, evaluation, governance, GTM, and pricing are no longer separable workstreams. Treating them separately creates systems that demo well but fail when buyers ask about cost, reliability, rollout, and accountability.

Infrastructure note

Token pricing is a product design constraint now.

Input tokens, output tokens, cached context, tool calls, retries, eval passes, and latency tiers all shape the cost of a customer workflow. The product team may think it is designing an assistant, but finance sees a variable cost engine. Pricing has to connect those views without making customers feel like they are buying raw compute.

Price the outcome or workflow when customer value is stable.
Meter usage when consumption changes cost in a way the customer understands.
Use credits or allowances when the system needs room for different task sizes.
Expose limits when abuse, long context, or retries can distort margin.

Outcome quality

Excellent AI is measured at the point of work, not at the model leaderboard.

A model can score well and still fail the product. The product promise should define what good means: a correct answer, a completed task, a resolved case, a safe recommendation, a clean handoff, or a decision that an expert can defend. That definition then drives evaluation, review, escalation, and pricing.

Build eval sets from real customer tasks, not only synthetic prompts.
Track failure types by business consequence.
Use human review where error cost is high or trust is still forming.
Keep an evidence trail for enterprise buyers and internal operators.

Agent orchestration

Most agent problems are really workflow boundary problems.

Agentic systems work when the tools, permissions, memory, stopping rules, and review points are explicit. Multi-agent architecture can be powerful, but it is expensive and harder to debug. The stronger move is to choose the simplest pattern that fits the work: one agent, a sequence, parallel review, a supervisor, or an evaluator loop.

Single agents are usually the starting point.
Sequential flows help when stages are known and auditability matters.
Parallel agents help when independent judgment reduces risk.
Supervisor agents help when specialists need routing and shared context.

Model tuning

Fine tuning is not the first answer. It is one answer.

The path depends on the failure. If the system lacks facts, improve retrieval. If it ignores instructions, tighten prompts and examples. If it needs repeatable domain behavior, consider tuning. If the workload is high volume, consider distillation, routing, or a smaller specialized model. The business question is which intervention changes the outcome enough to justify cost and maintenance.

Separate knowledge gaps from behavior gaps.
Audit the quality and rights of training data before tuning.
Measure latency, margin, and support load after any model change.
Keep the product promise narrower than the lab result.

GTM and field

The field needs proof before it needs more narrative.

Enterprise AI sales breaks when reps can describe the vision but cannot defend the workflow. Buyers want to know what happens when the agent is wrong, what data is used, who reviews the output, what the rollout costs, and how success will be measured. The story has to carry those details without sounding defensive.

Turn pilots into production criteria before the pilot starts.
Give sales language for limits, not only benefits.
Package implementation work honestly.
Connect partner and marketplace motion to a real offer.