01
Desktop GUI and CLI
Select a JPG, PNG, or WebP loading screen, run analysis in a background worker, preview the result, and export the HUD without replacing the command-line path.
Windows development alpha · local screenshot workflow
Turn a loading screen into clearer team context.
BlitzPredict analyzes a manually selected World of Tanks Blitz loading screen, recognizes both compact preferential and wider standard 7v7 rosters, resolves player identities conservatively through the official Wargaming API, and renders an exportable tactical HUD.
Player recognition and HUD rendering run locally. Automatic capture, a live overlay, a hosted backend, and public distribution are not part of the current build.
PLAYER_A0151.8% · 2.4k
49.6% · 1.9kPLAYER_B01
PLAYER_A0247.2% · 1.7k
52.4% · 4.1kPLAYER_B02
PLAYER_A0350.1% · 3.0k
48.8% · 2.2kPLAYER_B03
OCR_UNCERTAINABSTAINED
54.0% · 8.6kPLAYER_B04
PLAYER_A0553.3% · 5.2k
ABSTAINEDNO_EXACT_MATCH
PLAYER_A0648.9% · 2.6k
50.7% · 3.8kPLAYER_B06
PLAYER_A0750.4% · 3.3k
51.2% · 3.5kPLAYER_B07
Stage
Development alpha
Interfaces
Windows GUI + CLI
Input today
Manual full-screen image
Current focus
Measured OCR quality
Built today
A working local application, not a notebook-only concept.
The original experiment has been extracted into an importable Python package with reusable services, a command-line workflow, a Windows desktop interface, deterministic tests, and an internally validated portable build path.
02
Adaptive 7v7 geometry
A shared planner supports compact preferential and wider standard rosters, preserves source aspect ratio, and retains a contour fallback for unfamiliar screens.
03
Local nickname OCR
EasyOCR reads row crops produced by the same geometry plan used across production analysis and controlled development audits.
04
Conservative identity trust
Only a unique, case-insensitive exact nickname match contributes statistics. Corrected, fuzzy, saturated, duplicated, or ambiguous results abstain instead of silently misattributing a player.
05
Managed Wargaming access
Official public API calls use run-local caching, bounded retries, conservative pacing, batched account information, and account-ID keyed joins.
06
HUD and honest diagnostics
The renderer keeps failed or uncertain rows visible, reports the trusted sample size, and exposes aggregate progress without nicknames, IDs, paths, coordinates, or image content.
Current pipeline
One local flow with explicit trust boundaries.
Geometry and OCR happen locally. Player statistics are requested only for identities that pass the exact-match safety contract.
01 / INPUTFull screenshot
A manually selected battle loading screen.
02 / PLANRoster geometry
Normalize, identify the layout, and produce 14 row crops.
03 / READLocal OCR
Extract nickname candidates with deterministic preprocessing.
04 / TRUSTExact identity
Reject uncertain or ambiguous account resolution.
05 / FETCHPublic statistics
Use managed requests to the official Wargaming API.
06 / OUTPUTRender and export
Build the HUD, preview it, and save the result.
Local: source image, geometry, OCR, result construction, rendering.
Network: official Wargaming endpoints for trusted account and statistics requests.
Evidence before claims
Reproducible contracts around a still-improving OCR core.
BlitzPredict separates deterministic engineering checks from private development evidence. Passing infrastructure checks does not get presented as product accuracy.
What is measured now
The repository includes offline pipeline fixtures and controlled local audit runners. Each tool states what it proves and what remains outside its scope.
Cross-platform Python CIWindows and Ubuntu across Python 3.11, 3.12, and 3.13.
Active
Deterministic synthetic benchmarkReal service wiring with scripted OCR, invented accounts, a fake API, and a socket guard.
Committed
Private geometry auditAggregate-only checks over both roster cohorts with private inputs kept outside Git.
Available
Private real-OCR auditLocal model weights, downloads disabled, network blocked, aggregate-only output.
Available
Windows portable validationLocked CPU build, archive extraction, structural checks, and a native OCR import probe.
Internal
Architecture boundary
The local core exists. Production delivery remains a separate decision.
The current application does not depend on a BlitzPredict cloud backend. A hosted service should be introduced only if a controlled pilot shows that it is necessary.
Windows screenshot analysis
A Python application with shared services for GUI, CLI, benchmarks, audits, and the portable build.
- Manual local file selection and local image processing.
- Adaptive and fallback roster geometry.
- EasyOCR model stored on the user's machine.
- Direct official Wargaming API requests for trusted identities.
- Local HUD preview and image export.
Controlled distribution and optional services
These items are roadmap candidates, not deployed product capabilities.
- Signed Windows distribution with licensing and notices complete.
- A documented automatic capture target and permission model.
- A small private pilot with cost and reliability measurements.
- An optional authenticated backend, cache, and monitoring only when needed.
- Publisher-policy review before any live-use release.
Roadmap
Improve the evidence, then expand the product surface.
Extract and harden the application core
Package, GUI, CLI, safe matching, managed API access, adaptive geometry, CI, portable build validation, and aggregate diagnostics.
Run one bounded OCR-quality iteration
Compare deterministic crop preprocessing against the same private development baseline and keep the change only if aggregate guardrails improve.
Measure native latency and choose capture scope
Record controlled cold and warm Windows runs, then document the first automatic-capture platform, permissions, trigger, and manual fallback.
Prepare responsible distribution
Complete licensing and signing work, run a limited pilot, measure reliability and cost, and review Wargaming policy before a broader release.
Current limitations
Clear boundaries make the next milestone measurable.
Manual input
The current GUI opens a saved screenshot. It does not detect or capture the game window automatically.
No live overlay
The application renders an output image after analysis. It is not an in-game or always-on overlay.
OCR still under development
The private baseline selects the next experiment; it is not a release threshold or independent holdout.
Internal portable build
The Windows package path is validated in CI but remains unsigned and is not publicly distributed.
No deployed backend
There is no production BlitzPredict cloud, subscription service, user account system, or mobile application.
Policy review remains
Any live-use release must still be checked against Wargaming policy and applicable platform rules.
Responsible scope
Useful context without hidden access to the game.
BlitzPredict is designed around visible loading-screen information, public player statistics, conservative identity handling, and explicit evidence limits.
- No game-client code injection, process memory reading, packet interception, or gameplay automation.
- Only unambiguous exact account matches may contribute statistics to the HUD and team sample.
- Committed benchmark data is synthetic; private development screenshots and audit outputs stay outside Git and CI.
- Audit output is aggregate-only and excludes nicknames, screenshots, crops, paths, account IDs, coordinates, and per-row results.
Build the next evidence milestone
BlitzPredict is ready for focused development, not inflated promises.
The immediate work is concrete: improve OCR against fixed privacy guardrails, measure native performance, choose one capture target, and prepare a responsible private pilot. For technical, product, or pilot discussions, contact Karol.
BlitzPredict is an independent project. World of Tanks Blitz and related marks are trademarks of Wargaming. This project is not affiliated with or endorsed by Wargaming.