1 of 10

Winner • Singularis

Hackaging.ai

2025

Team Vincent & Dmytro

Scientific Knowledge Graph Engine

2 of 10

Meet the team

Dmytro Lozovyi

Engineer

Vincent Ilinicz

Engineer

3 of 10

Read the text

Takes time and requires search and lots of attention

​

Existing Datasets

Lack of article analysis, broken links.

​

​

Use GPT

Costly and not deterministic.

When it will loose context next time?

​

The problem

​

It’s not so easy to surf through scientific papers

In 21 century we still have

no unified knowledge

and research system

4 of 10

The solution

Deterministic text classification and visual clarity

  • Extract knowledge atoms with traceable NLP rules
  • Display extracted data in an easy visual way - graph
  • Preserve provenance and atom-to-article reference

We’ve built an extensible IMRAD-styled articles parser and classifier as all-in-one solution: load, extract, link, display.

5 of 10

1 Extract Text

3 Link and display

Our Approach: 3 steps

2 Label sentences

Custom pymupdf or full-fledged docling are alternatives, but requires either much more resources or development time

Connect atoms with similarity, sliding window and strict transition rules:

Analysis→Conclusion

​

Transparent and accessible

Display in browser.

OS-aware, easy to use

even locally.�Zoom-in, zoom-out, �see highlights in original PDF�

We use GROBID pdf parser as stable and balanced solution.�Provenance and structure.

​

Used for years

Rule-based pattern engine.

SpaCy, lemmas, dependent matching, polarity

​

Extensible and precise

Hand crafted deterministic scoring, label concurention, IMRAD-based weights, confidence calculation,

​

​

6 of 10

Labeling and how it works

RESULT_LIKE: 1�RESULT_DEP: 1.2

IMRAD: RESULT => Result += 1

LACK OF CITE => FACT -= 1

“We also showed that L. fermentum influences only grade C1”

It’s “RESULT”!

FACT_LIKE: 1

FACT_DEP: 1.2

Final decision

Post-Filters, Tie Breaking

Rule Hits

Input sentence

7 of 10

Graph and Provenance

See full Graph �

Connected �article atoms

Metadata for each node

​

PDF article linkage

8 of 10

Matching Rules

Graph Edges

LLM Feedback

Next steps and improvements

Make matching rules more smart:

  • section aware
  • FoS aware
  • weighted
  • demotion

​

Semantic linking across whole article.

​

Contextual.

​

​

Ask LLM to generate new rules per problematic article with low confidence score.

Stability via golden reference set.

​

9 of 10

Any questions?

Ask away!

​

​

Vincent: @vilinicz

Dmytro: @lozovoi_dn

10 of 10

Thank you