Eleonora Drykina is a design engineer whose work spans model-agency software, expressive browser graphics, augmented reality and AI interfaces. Her portfolio includes both visual experiments and tools for practical tasks: managing agency content, navigating a festival and finding answers in payment rulebooks. She collaborated with Nina Žižakić on the payment-rule assistant at In The Pocket.
From fashion to interactive software
Eleonora’s fashion background includes responsibility for direct bookings at Genom Management. She also built the agency’s website using Astro and Strapi, allowing agents to manage its content. The project brought her software work into the operations of an industry she already knew.
Her DEVINE studies from 2022 to 2025 combined design and programming. Her portfolio shows several directions within that practice:
Liminal Russia: A personal exploration of WebGL and WebGPU for self-expression and processing experience.
Transfo Intiem: A 3D website and augmented-reality experience designed to help festival visitors navigate the site and its schedule.
Scapino: A collaborative school project pairing an immersive installation with a companion website. The installation combined a p5.js shader-based video filter with Arduino and NFC interactions; the website incorporated 3D models.
Preserving evidence in AI answers
Eleonora and Nina’s payment-rule assistant addressed a different navigation problem: a question might require one table row, a whole section or information spread across several rulebooks. Built for an anonymous Belgian payments client, the system worked with about 270 PDFs containing 6,000 pages.
Their indexing pipeline preserved section paths and chunk order so the assistant could reconstruct larger passages from retrieved fragments. It also converted tables into CSV while retaining the relationships between headings and values. These choices mattered because retrieving an isolated value without its heading or surrounding rules could strip away the information needed to interpret it. The agent could expand its search from initial results to sections, full documents and multiple rulebooks, with limits to prevent oversized documents from overwhelming its context.
The interface had to accommodate a median answer time of 38 seconds. Streaming Markdown separately from a completed source card let users begin reading before the full response structure was ready. Evaluation also required looking beyond a score: Eleonora and Nina described cases where an automated judge’s apparent regression did not hold up when checked against the source documents. Their account connects the assistant’s retrieval, presentation and evaluation through concrete decisions about how users receive and inspect evidence, while retaining joint attribution for the work.
Eleonora Drykina and Nina Žižakić explain how a payment-rulebook assistant preserves table meaning, reconstructs larger passages and widens retrieval—then makes a 38-second answer usable and its evaluation open to challenge.
Non-overlapping chunks, section paths and chunk indexes let the agent reconstruct larger passages instead of relying only on isolated search hits.
Hybrid dense and BM25 retrieval raised the reported retrieval hit rate from 92% to 97% on 63 questions; this was not an end-to-end answer accuracy result.
Evaluation changes need investigation: judge decisions can vary between runs, and an incorrect rubric can penalize a fact that is present in the source.
▶ Watch ↗