Input
A manually supplied battle loading-screen screenshot containing two seven-player teams.
Working proof of concept · manual screenshot input
From loading screen to readable player context.
BlitzPredict recognizes the 14 players visible on a World of Tanks Blitz loading-screen screenshot, corrects OCR errors, retrieves public statistics through the official Wargaming API, and renders a tactical HUD with Win Rate, average damage, and team averages.
Current implementation: Python notebook in Google Colab. The production client and cloud backend are planned, not yet deployed.
Stage
Working proof of concept
Current stack
Python · EasyOCR · OpenCV
Input today
Manually supplied screenshot
Next milestone
Benchmark and standalone client
Working today
A real input, a real pipeline, and a real generated result.
The current proof of concept validates the technical core. It does not yet claim automatic screen capture, a production backend, live overlay delivery, or commercial readiness.
Generated output
OpenCV masks the original nickname regions and renders returned WR, average damage, and team averages.
Current capabilities
The prototype already proves four critical components.
01
Targeted OCR
EasyOCR processes the two player-list regions instead of scanning the entire screen without structure.
02
Custom correction
A project-specific correction and cascading lookup system improves account matching when OCR misreads nicknames.
03
Official API data
The prototype resolves player accounts and retrieves public Win Rate and average-damage statistics through Wargaming API endpoints.
04
HUD generation
OpenCV renders a usable result directly onto the source image and calculates average team Win Rate.
Current pipeline
Clear enough to test, small enough to improve.
01 / INPUTScreenshot
Manual proof-of-concept input.
02 / CROPPlayer regions
Two team areas and fourteen rows.
03 / OCRNickname reading
EasyOCR extracts candidate text.
04 / CORRECTAccount matching
Custom correction and cascading lookup.
05 / APIPublic statistics
Wargaming account and performance data.
06 / RENDERTactical HUD
OpenCV output and team averages.
Planned production architecture
A focused cloud path, clearly separated from the current PoC.
The first production pilot should stay deliberately small. More complex streaming and historical analytics services can be added only when usage and benchmarks justify them.
01 / CLIENT
Android or Windows client
Capture workflow, local preprocessing, user controls, and result presentation.
02 / API
Cloud Run backend
A containerized Python/FastAPI service for authenticated requests, account resolution, and analytics logic.
03 / DATA
Cache and managed storage
Reduce repeated external API requests and store only data required for the authorized workflow.
04 / OPERATE
Logging and monitoring
Measure failures, latency, cost, API behavior, and pilot reliability before broader release.
This architecture is a roadmap for turning the validated notebook pipeline into a controlled, measurable pilot. It is not presented as an already running production system.
Planned validation metrics
The next version will replace placeholders with documented measurements.
No benchmark numbers are invented here. Results will be published after a repeatable test set, methodology, and sample size are defined.
Nickname accuracy
Exact-match success across varied screens, aspect ratios, clan tags, underscores, and long names.
TBD after test setAccount resolution
How often OCR candidates resolve to the correct Wargaming account after correction and cascading lookup.
TBD after test setEnd-to-end latency
Time from available screenshot to completed statistics and rendered HUD.
TBD: p50 · p95Pilot unit economics
Estimated cloud cost per processed battle and per active user at controlled pilot volume.
TBD after cloud pilotExecution roadmap
Build the evidence first, then scale the product.
Benchmark and harden the PoC
Build a documented test set, measure OCR/account-resolution accuracy, instrument timing, and catalogue failure cases.
Build the standalone client and controlled backend
Develop the Android/Windows workflow, authenticated API, cache, monitoring, and a limited private pilot.
Validate product demand and repeatable scale
Test subscriptions or team plans, verify cost assumptions, expand analytics, and complete publisher-specific compliance review.
Commercial direction and responsible scope
An ambitious future, without rewriting the present.
Planned business model
- Subscription access for individual advanced players.
- Clan and team plans with shared analytical tools.
- Controlled pilot access for coaches, tournament teams, or analytical partners.
- Future modules may include session analytics, prediction, and coaching support after sufficient data and validation.
Product boundaries
- Publicly visible on-screen information and official public API data.
- No game-client code injection, memory reading, packet interception, or gameplay automation.
- Data minimization and no raw-image retention unless explicitly required and disclosed.
- Live-release scope remains subject to Wargaming policy review and applicable platform rules.
Founder
Product direction and technical ownership in one place.
Karol
Founder & Developer · Koksu Company
Karol built and tested the current Python proof of concept with AI-assisted development. He is responsible for the product concept, World of Tanks Blitz domain knowledge, OCR correction logic, API workflow, technical decisions, testing, and the path from notebook prototype to a commercial application.
Pilot and infrastructure support
Turn the validated core into a measurable product pilot.
Cloud resources would be used to package the backend, secure external API access, introduce caching and observability, benchmark performance, and support a limited pilot. The immediate goal is not unspecified scale. It is reliable evidence.
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.