I study how technological and innovation systems (notably for blockchain & AI) reshape markets, platforms, and information access —
using causal methods to measure what changes and mechanism design to think
about what should. Current focus: the microeconomics of artificial agents —
economics for decision-makers whose internal mechanisms can be observed and audited.
Pre-doctoral researcher, previously at IMT Business School and Oxford
Saïd Business School. UC Berkeley & Sciences Po alumnus. Ex tech management consultant at Wavestone.
Building toward a PhD in economics and an entrepreneurial journey.
Do Structural Models Recover Mechanisms? Evidence from Neural Agents with Observable Internals
[working paper · 2026]
Paper #1 of the Microeconomics of Artificial Agents program
When a structural model is estimated on an agent's behavior, do the recovered
parameters capture the agent's actual decision mechanism? With human data this is
untestable; with neural agents the mechanism is fully observable. In solved dynamic
discrete choice environments (Rust 1987), agents built on named bounded-rationality
primitives show that in-class inattention is benignly absorbed into as-if parameters —
counterfactuals survive — while out-of-class categorical cognition fails silently:
2–3% behavioral gaps yield badly biased structural estimates and 2× wrong policy
predictions, with immaculate likelihoods. Linear probes detect none of this, but
causal conformity audits of internal mechanisms rank-predict counterfactual failure
ex ante (Spearman 0.87 / 0.80, including a pre-registered experiment).
When to Clean the Machine: A Rust Model of Codebase Rationalization in AI-Native Startups
[working paper · 2026]
Adapts Rust (1987)'s bus engine replacement model to the software context, treating
codebase rationalization — the decision to rewrite fragmented, AI-generated
"vibe-coded" infrastructure — as a lumpy, partially irreversible investment under
uncertainty. AI shifts both structural objects at once: technical debt accumulates
faster per unit of feature output, and AI agents make large rewrites cheaper —
predicting rationalizations that become more frequent and less lumpy after AI
adoption. Taken to a 14-repository open-source panel (2023–2025), with AI-tool
release timing interacted with repository-level exposure as the identifying
instrument.
The Jevons Paradox in LLMs: Do Efficiency Gains Offset Token Generation Growth?
[working paper · 2026]
with Pierre Noro (Sciences Po) & Alan Seroul (ENS Paris Saclay)
Investigates whether growth in token generation per query by modern LLMs offsets,
partially or totally, the energy savings achieved through architectural and hardware
improvements — applying the Jevons paradox framework to AI compute economics.
Cryptocurrency Adoption and Payment Systems in Developing Economies
[course paper · 2024]
with Thomas Noel (Paris School of Economics) · UC Berkeley INFO 134/234
Builds a producer-theory model of Bitcoin adoption for SMBs, incorporating
network effects and exchange-rate volatility in economies considering crypto
as legal tender. Extended from a graduate-level IT economics seminar.
co-founder · Canopy Fellow at Founders Inc. ·
Website ↗
French SMEs leave operational surplus idle at 0% while short-term rates sit above 3% —
a firm with €150K idle forgoes thousands of euros a year, knows it, and never acts,
because the effort-to-reward ratio is unfavorable. YIELD removes the decision: it
connects the company's bank and accounting data (open banking, Pennylane), forecasts
cash flow to compute a safe-to-invest surplus, and sweeps it into a regulated
money-market fund (Spiko, AMF-authorized) with a weekly one-click approval. The
distribution is the economics insight: sold through accountants as a recurring revenue
stream for their practice, making the expert-comptable the distribution node,
supervisor, and trust anchor. Grew from ETHGlobal hackathon wins (SF & Cannes)
into a live product in France.
CompasSup & Clerko
[EdTech · 2022–2025]
COO & co-founder · Sciences Po incubator ·
Website ↗
Applied an economics-of-information framing to higher-education orientation in
France — where students systematically under-sort into programs due to information
asymmetry. Built from a free Parcoursup matching app into an AI-native platform
automating university administrative workflows. 7k users, 12M social views,
partnerships with public institutions.
Then pivoted to making an AI-native ERP software to enable Higher Education Institutions
to provide decent services and work processes to their students and collaborators ; got contracts
and advanced leads with Europe top schools, but left the space as human-friendly innovation was hated in
such organizations.
Cross-chain lending yield optimizer using MCP-powered multi-agent workflows.
Designed around household purchasing-power and inflation framing in emerging
economies — the mechanism design question being: can DeFi liquidity markets
serve as a hedge when local monetary policy fails? Best consumer app award
and best cross-chain DeFi use case at ETHGlobal San Francisco and Cannes.
