I build AI tools that make sense of the world, turning raw information, events, and news into something structured, useful, and honest.
I use AI to turn ideas into working software. SEFIN is a full AI research system I designed and shipped without an engineering team: the product concept, the analytical methodology, the system architecture, and the interface. My goal is to build sophisticated vertical AI tools that do not just retrieve information, but reason about it.
Demonstrated: live systems belowI build systems that reason professionally. The tools I build enforce structured thinking at the prompt layer: claim classification, evidence grading, explicit confidence ceilings, source standards. The default failure mode of AI tools is confident-sounding output regardless of evidence quality. I design mine to output based on rigorous analytical discipline, the same standards used in intelligence and investment research.
I work across cultures, languages, and current events. English native, Hebrew working proficiency, Dutch conversational. I have lived and worked in the US and Israel, have a genuine interest in international affairs and how the world fits together, and I build tools that reflect that depth.
Takes a geopolitical event (a strike, a sanctions package, a procurement decision, a shipping disruption) and produces a structured intelligence product: confirmed facts, transmission channels, exposed sectors and companies, confidence levels, and signals to monitor. Deliberately produces research leads, not recommendations.
An evaluation system that scores SEFIN.AI's analytical calls over time. Each assessment is logged at publication and graded against outcomes, with the heaviest weighting on non-obvious, upstream calls. Built on the principle that analysis without accountability is opinion.
Nine claim types, graded L1–L4 by evidence strength. The claim type sets the confidence ceiling for everything downstream.
No mechanism, no market claim. Every analytical link must trace a named causal pathway from event to P&L.
Four-tier source grading with recency rules. Ten weak sources never equal one strong one. AI summaries are not sources.
Confidence and likelihood are different things, kept separate per ICD 203. Confidence is earned by evidence, never by polish.
Every research lead must name the specific evidence gap and the exact disclosure that would upgrade it to a confirmed claim. Ambiguity is not a placeholder. It is a tracked variable.
Partial evidence yields labeled research leads, not silence and not false certainty. The output type must match the evidence state at every rung of the ladder.
Six versioned methodology documents adapted from ODNI ICD 203, the CIA Tradecraft Primer, CFA Institute Standards of Practice, Shell scenario methodology, and Damodaran country-risk research. Available on request.
Advised international investors on Israeli market opportunities, working in English and Hebrew.
Coursework: Business Strategy, Startup Marketing, Information Systems, Business Leadership · Board member and editor, YU Political Action Club: wrote on geopolitics, security policy, and economic affairs