Daniel Alami
AI product, adversarial evaluation, formal reasoning, and enterprise governance.
I work on AI products and public research where evaluation, verifiable reasoning, and governance are part of the system itself.
Most of my work 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.
I am a Senior Product Manager - Technical at Amazon Web Services. My independent public research is separate from my employment and does not disclose confidential or proprietary AWS information.
I earned my MBA from Harvard Business School in May 2026. Earlier in my career, I worked across AI, cloud, digital banking, and transformation programs at Microsoft, Santander, Gartner, and IBM.
Amazon Web Services. Harvard Business School MBA, 2026. Independent AI evaluation and governance research.
Public projects
These projects examine one recurring problem: telling when an AI workflow is genuinely improving rather than merely learning to pass its own checks.
- 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
- BlindSpot-Bench tests how dependence among evidence sources affects a supervisor's reconstruction of a changing system. Its current public study separates reporter count from independent evidence count and measures the cost of checking the system outside the reporting channel. Code Write-up
- Cognitive Firm is an operating model and runtime for AI work where generation, evaluation, memory, and authority remain separated so unsupported outputs cannot advance without recorded evidence and review. GitHub
- nonmathlib4 is a Lean 4 library of machine-checked propositions and counterexamples across several non-mathematical domains, with explicit attention to whether a formal claim preserves the original requirement. GitHub
Research program: from requirements to AI assurance
My research program examines how abstract properties of complex systems can be translated into explicit requirements, measurable indicators, and inspectable evidence. In 2015, I operationalized software-ecosystem health through observable measures. In 2017, I studied how security requirements could be expressed through socio-technical models and reinforced through feedback embedded in the modeling workflow. Related work examined structured, reusable institutional data.
My current research brings this methodological foundation to AI assurance, where an additional problem arises: the system producing a test, argument, or evidentiary claim may also influence the process used to evaluate it. I study how to detect that failure through adversarial execution, separated generation and evaluation, machine-checkable reasoning, rules that stop unsupported outputs, and preserved evidence.
The research domains and technical methods have evolved substantially. The continuing thread is the operationalization of difficult-to-observe system properties so that consequential claims can be tested, reviewed, and governed rather than accepted through self-certification.
This program is closely related to the testing, evaluation, verification, and validation practices described in the NIST AI Risk Management Framework, including documented test sets, repeatable measurement, independent assessment, deployment-relevant evaluation, and explicit limits on generalizability. This reference describes standards relevance; it does not imply NIST review or endorsement of my work.
Core AI evaluation research
These 2026 preprints report formative studies and have not undergone peer review. I am consolidating them into a larger held-out evaluation of specification gaming and independent checks.
- Specification gaming in LLM-generated code: detecting cognitive camouflage by adversarial execution. Reports nine specification-gaming strategies from 453 adversarial debate logs across six domains and tests adversarial counter-execution against holistic model judging. Preprint, 2026. SSRN GitHub
- Hardening an adversarial evaluator with mined constraints and contract-gated recursive improvement. Tests rules that block unsupported outputs, mined failure constraints, and typed review contracts across six evaluator-hardening stages. Preprint, 2026. SSRN GitHub
Exploratory and superseded drafts remain archived for provenance but are not presented here as current research results.
Earlier peer-reviewed work
- A Gamified Tutorial for Learning About Security Requirements Engineering. Daniel Alami and Fabiano Dalpiaz. IEEE International Requirements Engineering Conference, 2017.
- Relating Health to Platform Success: Exploring Three E-commerce Ecosystems. Daniel Alami, Maria Rodriguez, and Slinger Jansen. Workshop on Ecosystem Architectures, 2015.
- 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 record marks the earlier stages of the same methodological progression: explicit requirements, measurable system properties, inspectable evidence, and tools that embed verification into the working process.