javohir.engineer logo
javohir.engineerSoftware Engineer
Projects/ENGLISHAI
CASE STUDY 01BACKEND · AI

EnglishAI

Adaptive English platform with speech assessment, placement testing and real-time pronunciation scoring.

ROLEBackend Developer
TIMELINEJul 2026 — Present
STACKASP.NET Core · PostgreSQL · Redis
STATUSLive · 700+ users
ENGLISHAI / ADAPTIVE ENGLISH PLATFORM
700+ACTIVE LEARNERS

Placement testing, pronunciation scoring, and lesson progress on a unified .NET backend.

EnglishAI Platform Interface
AZURE SPEECH · PRONUNCIATION
0:07 / 0:07CACHED · <120 ms
THE BRIEF

Adaptive English testing, lesson progression, and real-time pronunciation scoring.

MY ROLE

ASP.NET Core backend architecture, Azure Speech integration, and Redis caching.

KEY GOAL

Low-latency speech scoring under 120ms and resilient test session recovery.

01 — THE PROBLEM

The platform worked.
The onboarding didn't.

Telemetry revealed that legitimate mobile users were being dropped due to false-positive anti-cheat flags during placement tests. We refined the validation logic and enabled session recovery so students never lose test progress.

21 / 27
drop-offs traced to false mobile anti-cheat triggersOnboarding completion rose 78% after the fix.
02 — WHAT I BUILT

Three pieces, one flow.

Every decision below is in production.
01

Core API Architecture

ASP.NET Core REST endpoints for placement tests and lesson tracking, backed by PostgreSQL and Redis read caching.

Clean architecture · EF Core · Redis
02

Speech Evaluation Pipeline

Azure Speech STT and pronunciation assessment with MD5-hashed Redis caching, cutting duplicate third-party costs.

STT · scoring · MD5 caching
03

Resilient Session Recovery

State persistence and idempotent resume logic allowing students to resume interrupted tests without losing progress.

State persistence · zero data loss
03 — ARCHITECTURE

How the pieces fit.

Simplified request path
CLIENTweb · telegram mini app
ASP.NET CORE APIrest · auth · validation
POSTGRESQLsource of truth
REDIScache · pub/sub
AZURE SPEECHstt · assessment
BACKGROUND JOBShangfire · recovery
audio key = MD5(text + voice) → hit rate ↑, spend ↓ 35%Sub-120ms roundtrip on cached audio
04 — RESULTS

Measured after shipping.

Production telemetry · team dashboards
700+
ACTIVE LEARNERSon the adaptive platform
+78%
ONBOARDINGcompletion after recovery
−35%
SPEECH COSTvia Redis audio cache
<120 ms
REPEAT LATENCYfor cached pronunciation