I study how machines read imperfect traces.

My work is on AI for difficult documents: receipts and insurance claims, epidemic reports, news and social media, historical newspapers and archives. I build methods to extract people, places, events and relations from them, and I measure what models miss, invent or flatten along the way.

La Roche1lle, l14 juillet 189O0. — Le Consceil municipal s’est réuni à l’Hôtel de Vi11lle.

  1. Meeting of the city councilevent
  2. La Rochelleplace
  3. 1890-07-14date
  4. Conseil municipalorganisation
  5. Hôtel de Villeplace

Épicer1ie du Port · 14/03/2023 · café 5,10 · pain 2,40 · TOTAL 9,50≠ 7,50

  1. café + pain = 7,50, but the total says 9,50altered?
  2. Épicerie du Portcompany
  3. 2023-03-14date
  4. café, 5,10product
  5. pain, 2,40product

12 March: 14 suspected cases of cho1lera reported in Kisumu, health officials said.

  1. Cholera outbreakevent
  2. choleradisease
  3. Kisumuplace
  4. 14suspected cases
  5. 12 Marchdate

omg the river is ov3er the bridge on Rue du Port!! pplpeople stuck on roofs, pls send boats #flood

  1. Flood, people strandedevent
  2. Rue du Portplace
  3. people on roofsaffected
  4. boatsneed
  5. #floodtopic

Six threads I keep returning to

Semantic extraction

Named entities, entity linking, relations and events: the small structures that make collections searchable.

[La Rochelle] place

Applied document intelligence

Document fraud, misinformation, epidemic surveillance, and open LLMs built to hold up under real constraints.

a receipt: is the total real?

Noisy documents

OCR and HTR errors, degraded text, and how noise changes downstream meaning, not just character accuracy.

Vi11e → Ville

Multimodality

Text, images, layout and visual signals modelled together, rather than treating everything as plain text.

text + image + layout

Historical AI

Temporal modelling and multilingual archives, for language whose names, borders and meanings keep shifting.

“ie” for “je”, 1630s

Cultural heritage

Digital epigraphy, Armenian and Ukrainian inscriptions, structured vocabularies, and contested memory.

։ the Armenian full stop, not a colon

Selected papers

    Noisy documents Semantic extraction Historical AI Cultural heritage Multimodality Applied document intelligence
    All 85 publications by theme and year.
    The same 85 publications by theme.
    The same 85 publications by type.
    Full list: 85 publications →

    Notes and essays

    All writing →

    Mostly serious work, with a sense of humour

    I care about what AI extracts from documents, and also about what it misses, invents, flattens, or quietly ruins while sounding confident.

    I come from both computer science and visual practice, so I tend to look at documents as text, images, layouts and traces all at once. My PhD was on neural methods for event extraction; since then I have worked on historical newspapers, epigraphy, document fraud, epidemic surveillance and open LLMs.

    Outside research: photography, handmade things, gardening with variable success, mostly non-fantasy books, metal festivals and camping.