Find the ground before anyone else knows it's there.

KazRock Global fuses a million declassified Soviet geological reports, live Sentinel satellite ground-truth, and six national licence cadastres into one calibrated number: the probability that a piece of unclaimed ground holds an economic deposit.

Soviet reports in scope
0
Jurisdictions
0
KZ application window
10 working days
Doc → ranked target
10 min
SCROLL
01The asymmetry

Three signals nobody else holds at once.

Each leg is public or semi-public on its own. The moat is the join, and the speed of acting on it.

01

The Soviet archive

From 1950–1991 the USSR systematically mapped all of Central Asia at 1:200 000 and 1:50 000: drillholes, geophysics, GKZ-classified resource estimates. Roughly a million scanned, Russian-language reports, inaccessible to Western explorers. Industrial OCR + LLM extraction collapses that barrier.

OCR · ru → structured
02

Satellite ground-truth

Sentinel-2 SWIR alteration mapping and Sentinel-1 InSAR independently corroborate or invalidate the archive at zero marginal cost per location. A Soviet P1 estimate and a strong iron-oxide + clay signal at the same coordinate is materially higher-conviction than either alone.

S2 · S1 · ASTER
03

First-come licensing

Kazakhstan and its neighbours issue solid-mineral exploration licences on a first-application basis, with statutory decision windows measured in working days. Speed of identification is the moat. The first correct application wins the ground.

6 cadastres · live
Archive
Satellite
Licence
KRG Target Score
02The compression

One calibrated number. Six measured features.

Not a vibes-based ranking, but a calibrated probability that drilling confirms or exceeds the historical estimate, recomputed every time new evidence lands.

0 P(discovery) · calibrated HIGH CONVICTION

Every score decomposes into six 0–10 sub-scores. Two clicks return the full evidence chain: the document, the satellite acquisition, the calculation.

  • AES Archive Evidence 9.1
  • SAS Satellite Alignment 8.4
  • CDS Commodity Demand 9.6
  • LAS Licence Availability 8.5
  • GAS Geological Analogue 7.8
  • LS Logistics 8.0
PHASE 1 Heuristic prior

Cold start. A weighted blend of the six features while the field record is thin.

PHASE 2 Calibrating

At 30+ drilling outcomes, a logistic model fits. Brier & ECE tracked on every result.

PHASE 3 Mature

At 200+ outcomes, a hierarchical model per terrane & commodity, with SHAP & conformal intervals.

Every drill hole KRG sinks recalibrates the model. No competitor can replicate it without the same field investment. The score compounds into IP.
03The operating system

From a scanned report to a signed deal.

Six surfaces, one boundary. Target lists, application drafts and royalty positions never leave the platform except through a watermarked data room.

/targets

Targets

A ranked, calibrated working environment: map left, target list right. Filter by commodity, score, licence availability, score-change direction.

/cadastre

Cadastre

Six national licence registries, snapshot-diffed daily. Lapse alerts at T-180/90/30/0 and competitor-activity alerts within 25 km of any live prospect.

/applications

Applications

Pre-filled, jurisdiction-aware licence bundles. The KZ set renders all six statutory documents from templates, geologist- and legal-signed.

/portfolio

Portfolio

Held-licence management with work-programme compliance tracking and offline-first field capture. Every sample logged re-scores its prospect.

/dealroom

Dealroom

Royalty, JV and sale structures. Investor-grade Technical Packages with Competent-Person attestation and PAdES digital signatures.

/desk

Desk

The AI analyst surface. Quick screens, deep dives and risk briefs. Streamed, grounded, every claim cited back to a source section.

04The Prospector

A continuous engine, not a tool.

Evidence lands; prospects emerge ranked. A 50-page Russian PDF becomes a scored target in under ten minutes. Coordinates rigorously converted, provenance stamped on every field.

  1. 1
    Ingest & OCRPaddleOCR (ru) + Tesseract fallback, 300 DPI, deskew, quality gate.
  2. 2
    ExtractClaude, prompt-cached, into a strict provenance-stamped schema.
  3. 3
    ConvertSK-42 / Pulkovo-1942 → WGS-84 via pyproj, with an error budget in metres.
  4. 4
    LinkFellegi–Sunter dedup against existing prospects; human review band.
  5. 5
    CorroborateSentinel-2 indices + cadastre intersect, async on the worker fleet.
  6. 6
    ScoreSix features → calibrated Target Score, with a full audit event.
05The territory

Six jurisdictions, built in from day one.

Each carries a distinct archive language, coordinate system, classifier and regulator. Jurisdiction is a column on every prospect, document and licence.

  • Cu Copper
  • Au Gold
  • Li Lithium
  • REE Rare earths
  • U Uranium
  • Mo Molybdenum

Auditability over cleverness

Every score, coordinate and estimate traces to a document, an acquisition or a named human. “Why does this score 87?” returns a complete chain.

Calibration over rankings

The score is a probability we hold ourselves to: Brier and ECE tracked on every drilling outcome, recalibrated as the record fills.

Provenance is a column

Source document, page, bounding box, extractor and confidence are set on every field. Queryable, not a comment.

Speed is a feature

Document to ranked target in ten minutes. Cadastre lapses surface within sixty seconds of detection. The first correct application wins.

Find the ground first. Drill the right metres. Sign the right paper. Compound the model with every hole.

The asymmetry is real.
The window is measured in days.

Single-tenant · internal operating system · not a SaaS