Baktiyar Ablaikhan
λ 522.7 nm

JerCase AI

Links scattered land-case documents to the right cadastral parcel, ranks the candidates and shows the evidence for each match.

Period
Jul 2026
Organisation
GovTech Camp 2026 (selection stage)
Role
Team baash, one of four
Prototype — the case workspace with a parcel history and top-3 candidates, and the spatial check
Prototype — the case workspace with a parcel history and top-3 candidates, and the spatial check
cases × candidates in the reproducible reference set
72 × 5
cases × candidates in the reproducible reference set
features in the reranker
21
features in the reranker

A specialist in a land department links each incoming document to a parcel by hand, across PDFs, scans, protocols and the cadastre. JerCase AI proposes candidates and shows why — source, reason, date — and leaves the final confirmation to the specialist.

How it works

Document text (OCR) → entity extraction → multilingual-e5-small embeddings → semantic top-K retrieval → CatBoost YetiRank reranking on 21 features → confidence and provenance, on a FastAPI/PostGIS backend with a Next.js workspace.

What is verified

Only a reproducibility check on a reference set of 72 cases × 5 candidates (360 pairs); the team’s own slide says it is not a production accuracy metric. Data were open, anonymised or simulated.

Figures

  1. 01InterfaceNext.js workspaceappeal · candidates · map · Decision Guard
  2. 02Application APIFastAPI + workflowingestion · reconstruction · review
  3. 03IntelligenceOCR · E5 · CatBoost · PostGISretrieval · ranking · spatial / temporal
  4. 04Evidenceprovenance layerdocument → entity → parcel → decision
  5. 05DataPostgreSQL · PostGIS · pgvectorsingle store for the case and geometries
Source adapters
  • eOtinish
  • EGKN
  • Documentolog
  • Lotus
  • Archive
Five layers from the case workspace down to a single PostgreSQL store.Original ↗
  1. 01Document textOCR / native text
  2. 02Entity extractioncadastre number · address · date
  3. 03E5 embeddingsmultilingual-e5-small
  4. 04Candidate retrievalsemantic top-K
  5. 05CatBoost rerankingYetiRank · 21 features
  6. 06Confidence + provenanceverifiable result
Hybrid ranking: text, time and geometry. The core runs without an LLM.Original ↗

Automated A person decides

  1. 01Uploaddocument
  2. 02OCRtext
  3. 03Entitiesnumber · address
  4. 04RetrievalTop-K
  5. 05Reranking21 features
  6. 06Confidencescore
  7. 07Checkmap · time
  8. 08Confirmationspecialist
From an uploaded document to a recommendation a specialist confirms.Original ↗

Evidence

  • 01What was verified, what is expected, and what a pilot would still have to measure

From the presentation

  • Flow from an incoming document through entity extraction and candidate ranking to a land parcel and its evidence (source, reason, date), with final confirmation left to the specialist.
    Solution overviewFlow from an incoming document through entity extraction and candidate ranking to a land parcel and its evidence (source, reason, date), with final confirmation left to the specialist.
  • Diagram of one land case drawing on scattered PDFs, scans, resolutions, protocols, tables, archives and cadastre records, leading to an incomplete picture.
    Problem: scattered sourcesDiagram of one land case drawing on scattered PDFs, scans, resolutions, protocols, tables, archives and cadastre records, leading to an incomplete picture.
  • Seven manual steps a land-department specialist performs to link a document to a parcel, with the resulting losses (time, missed records, no single picture).
    Current manual processSeven manual steps a land-department specialist performs to link a document to a parcel, with the resulting losses (time, missed records, no single picture).
  • Cover slide: "JerCase AI — intelligent linking of state documents and land parcels", team baash, GovTech Camp 2026 selection stage.
    CoverCover slide: "JerCase AI — intelligent linking of state documents and land parcels", team baash, GovTech Camp 2026 selection stage.