Daniel Alami

Daniel Alami

AI product, evaluation systems, and enterprise governance.

I work on AI products where evaluation, release gates, and governance are part of the system itself.

Most of it sits where product meets evaluation: the part of a system that decides whether something is actually ready to ship, and the review loops and controls around that decision.

For most of my career I have helped large institutions turn messy technology shifts into something they can actually operate.

I have worked across AI, cloud, digital banking, and transformation programs at Microsoft, Santander, Gartner, and IBM. I am currently completing my MBA at Harvard Business School.

Harvard Business School MBA, 2026. Previously Microsoft, Santander, Gartner, and IBM.

Public projects

These are built around one problem: telling when an AI workflow is genuinely improving rather than just learning to pass its own checks.

  1. ZTARE is an adversarial evaluation system for finding when model-generated tests, evidence, or arguments look rigorous while missing the thing they were supposed to check. GitHub
  2. Cognitive Firm is an operating-model sketch and runtime for AI work where generation, evaluation, memory, and promotion gates need to stay separated. GitHub
  3. BlindSpot-Bench measures the gap between what an overseer believes and what is actually true when the system being supervised can game the reporting channel. Code Write-up
  4. deai-mirror measures how far a piece of writing sits from how real people write in a chosen venue, then names the exact features that read machine-made, down to the over-used word. It compares against a real corpus instead of asking a model whether the text looks human, which is just the model grading itself. GitHub

How I think

The thread through all of it is supervision under pressure: a system learning to look good to whatever grades it without getting better at the actual task. The same pattern turns up in model benchmarks, in software tests, and in large organizations where the reporting quietly drifts from reality.

My instinct is to make the evaluation itself inspectable: keep the actor and the judge apart, leave the evidence visible, and push adversarial cases at the evaluator instead of trusting it.

Background

Before the AI work I published on security requirements, software ecosystems, and open data, all of it circling the same question of inspectability: how technical systems behave once people, platforms, and incentives are part of them.

The industry side gives me the other half of the problem: making new technology usable inside institutions that have to answer for it, not only demo it.

Selected AI Evaluation Work

Recent technical work behind the artifacts.

  1. Cognitive Camouflage: Specification Gaming in LLM-Generated Code Evades Holistic Evaluation but Not Adversarial Execution. Preprint, 2026. SSRN GitHub
  2. Hardening an Adversarial Evaluator with Mined Constraints and Contract-Gated Recursive Improvement. Preprint, 2026. SSRN GitHub
  3. The Limits of Audit for Self-Evaluating AI: Authority, Correlated Error, and the Price of Judgement. Preprint, 2026. GitHub
  4. The Cognitive Firm: A Multidivisional-Form Architecture for Recursive AI Governance. Preprint, 2026. SSRN GitHub

Peer-Reviewed Work

  1. A Gamified Tutorial for Learning About Security Requirements Engineering. Daniel Alami and Fabiano Dalpiaz. IEEE International Requirements Engineering Conference, 2017.
  2. Relating Health to Platform Success: Exploring Three E-commerce Ecosystems. Daniel Alami, Maria Rodriguez, and Slinger Jansen. Workshop on Ecosystem Architectures, 2015.
  3. Experiences from the Design and Development of an Institutional Linked Open Data Portal. Daniel Alami, Isaac Lera, Carlos Guerrero, and Carlos Juiz. TEM Journal, 2017.

This older work is why the current AI research leans so much on controls, incentives, and reusable evidence.