Anna Smirnova

AI research: synthetic populations, LLM measurement, evaluation.

Chief of AI at Foresenta. PhD candidate in economics at the University of Lausanne.

github ↗ papers cv (pdf) hi@aleph.science
// selected work
[ flagship ]

Foresenta↗ www.foresenta.com

Survey research on synthetic populations. I built the interview engine, where an LLM answers as each synthetic respondent, and the 1M-person UK population behind it. Our Swiss vote forecast was locked and published before the 27 September 2026 vote and came closer than every August poll.

forecast 35.1% · official 29.9%full write-up →
research
built

Synterr

Synthetic grammar errors, each tied to one of 48 rules, and evaluation rule by rule instead of one average score. First-author paper at BEA @ ACL 2026.

NLPGECevaluation
library + paper · BEA @ ACL 2026

PrismQL

Query language for patterns over time in conversations: this, then that, within ten minutes. Our AI Swarm Dynamics hackathon entry ran every question about agent swarms as a PrismQL query.

ANTLRPythonAI agents
language · open source

LLMs as measures of innovation

Every US LCD-industry patent classified by Gemini, fine-tuned on open-weight Gemma, and checked against real fab data: adjusted R² for welfare gains from ~0 to 12–14%.

GeminiGemmaecon
working paper · 2026

Lectern

Review app for speech-to-text transcripts: click a word to hear it, vim-style editing, agreement between reviewers.

human-in-the-loopspeech
web app

DEQN-JAX

Neural solver for high-dimensional economic models, in JAX. deqn-agent turns a paper into a trained, residual-checked solver.

JAXmacroagents
library · open source

proofmap

Web app that lets mathematicians map how a proof fits together, so the structure of the argument is visible at a glance. Built on React Flow.

React Flowmath
web app · live demo
// teaching

Data Science and Advanced Programming

A first programming course for the AI era, no prior experience assumed. Weekly workshops in browser-based cloud workspaces.

PythonHEC Lausanne
teaching assistant · UNIL

Advanced Data Analytics

Machine learning for economics and finance, from regression to reinforcement learning. Graded on a final project.

MLecon & finance
teaching assistant · UNIL
// papers
Patents as Indicators of Innovation: Large Language Models Meet Ground Truth in the LCD Industry
M. Igami, S. Kusaka, S. Scheidegger, A. Smirnova
Working paper, 2026
What Aggregate Scores Hide: Per-Rule Evaluation of Russian Grammatical Error Correction
A. Smirnova, A. Kopan, V. Makeev, G. Chernishev
BEA 2026 @ ACL · ties to Synterr
Query Processing and Optimization for a Custom Retrieval Language
Y. Kuzin, A. Smirnova et al.
Pan-DL @ COLING 2022 · ties to PrismQL
Situation-Based Multiparticipant Chat Summarization: a Concept, a Tool and a Collection
A. Smirnova et al.
ACL-IJCNLP SRW 2021
// about
Currently: Chief of AI, Foresenta · PhD candidate, University of Lausanne, advised by Simon Scheidegger · prev. Data Scientist, Universe Data
Education: BSc, MSc in computer science (Software and Administration of Information Systems), Faculty of Mathematics and Mechanics, Saint Petersburg State University · Swiss Program for Beginning Doctoral Students in Economics, Study Center Gerzensee, 2025