Writing
WHAT WE HAVELEARNED SHIPPINGTHIS.
- Pieces
- 16
- Topics
- 4
- Review cadence
- Quarterly
Latest
Most recent
Guardrails and control
9 min
Validating AI output before it reaches anyone
Four layers of validation between generation and delivery, ordered cheapest first, and what to do when each one fails.
Guardrails and control
10 min
Making AI decisions explainable, not just narrated
Why generated prose about a decision is not an explanation, and how to build one from the decision path itself.
Decision intelligence
11 min
How to rank options in a decision system
Option generation, normalisation, weights and aggregation, and why constraints must filter before ranking rather than score within it.
Agentic operations
10 min
What an AI agent should do when it fails
Five failure classes, retry limits that prevent incidents, and why partial completion is the hard problem.
Markets
9 min
Reward-to-risk, decided before you enter
Why a ratio computed after the position is open is a description rather than a constraint, and how position sizing reconciles stops with risk.
Markets
9 min
Analysing PSX stocks systematically
A repeatable evening process for covering the whole Pakistan Stock Exchange rather than the twenty names already on your screen.
Guardrails and control
9 min
What belongs in an AI audit log
A field-level specification for logging AI decisions, and the test that tells you whether your audit trail is actually complete.
Agentic operations
8 min
When not to use an AI agent
Six conditions that make a workflow a poor candidate for automation, and why deploying anyway industrialises the problem.
Decision intelligence
10 min
Dashboards, predictions and decision systems
Three things that get called the same thing, what each one actually gives you, and how to tell which one your problem needs.
Guardrails and control
9 min
Guardrails that are enforced, and guardrails that are merely requested
The difference between a constraint written into a prompt and one evaluated in code, and why only the second is a boundary.
Guardrails and control
10 min
Designing approval gates: which AI actions need a human signature
A tiered model for placing human approval where the cost of being wrong justifies the latency, and nowhere else.
Guardrails and control
10 min
How to detect model drift before your users do
Four proxy signals that move before output quality does, and how to set thresholds without guessing.
Agentic operations
9 min
How much autonomy should an AI system have
Autonomy derived from reversibility rather than from confidence, with the four factors that modify the baseline.
Decision intelligence
9 min
What a controlled AI pilot should measure
Why most pilots cannot fail, and the four measurements that make one decisive.
Markets
10 min
What a calibrated confidence score actually means
Calibration defined properly, how to check it, and why a flat reliability curve means the number should be removed.
Markets
11 min
When technical indicators disagree
Why conflict between indicators is the normal state, and four ways to resolve it consistently rather than by preference.
By topic
Grouped by what it is about
Guardrails and control
- Validating AI output before it reaches anyone
- Making AI decisions explainable, not just narrated
- What belongs in an AI audit log
- Guardrails that are enforced, and guardrails that are merely requested
- Designing approval gates: which AI actions need a human signature
- How to detect model drift before your users do
Decision intelligence
Agentic operations
Subscribe by feed: RSS
Bring us a real problem
If this is the kind of thinking you want applied to a decision you own, that is the conversation we want to have.