The Problem
Teaching English to Global business professionals is high-value work, but it doesn't scale. Every student needs personalized feedback, progress tracking, and live practice sessions. At any volume beyond a handful of students, the admin overhead consumes more time than the actual teaching.
The challenge: build a system that delivers the same quality of instruction to 50 students that a private tutor delivers to 5 — without hiring additional staff.
The Solution
CareerTalkLab is a standalone hybrid LMS at careertalklab.com. It uses a "Flipped Classroom" model: students complete web-based input (grammar, vocabulary, professional tasks) through a custom portal, reserving live Google Meet sessions exclusively for high-value output — speaking practice and real-time role-play.
The system runs two live tracks, with a third in planning:
- Executive Track — Business professionals (A1–B2) learning through task-based "Survival Sprints" framed around professional functions.
- Skills Lab — B1+ students working with news-based content (via an automated News-to-Lesson pipeline) for 2-hour intensive sessions.
- Exam Track (Roadmap) — IELTS/TOEFL candidates working through self-paced modules with AI-assisted feedback.
The Architecture
CareerTalkLab is a Next.js 16.1 project with its own domain, PWA support via Serwist, and a dedicated Supabase backend.
| Layer | Technology |
|---|---|
| Frontend | Next.js 16.1 (App Router), React 19, TypeScript |
| Styling | Tailwind CSS 4, shadcn/ui |
| Backend/Auth | Supabase (PostgreSQL, Auth, RLS, Edge Functions) |
| Content | JSON-native Lesson Engine with 8+ interactive block types |
| AI Feedback | Anthropic Claude API — drafts corrections on student writing |
| Pipeline | Python-based news ingestion (RSS → Gemini → Deploy) |
| Eval Suite | 4-tier automated testing (Unit, Pipeline, LLM-as-Judge, Trajectory) |
Key Technical Decisions
JSON-native Lesson Engine. Lessons are built on a robust JSON-native schema. A LessonRenderer handles specialized block types including <VocabularyCard />, <GrammarCallout />, <TrueFalse />, <Matching />, and <SubmissionForm />. This allows for programmatic content generation and rigorous schema validation.
Level-Calibrated Eval Suite. Content quality is enforced by a 4-tier eval harness that adapts per level:
- Unit Tests: Validates core logic and Zod schemas.
- Pipeline Tests: Checks for hallucinated vocabulary or grammar scope violations.
- LLM-as-Judge: A1 uses a 7-category rubric (GSE Assessment Framework + Errorless Teaching). A2 extends this to 8 categories, adding
chunking_qualityto evaluate the Proactive Exposure strategy introduced at that level. - Trajectory Tests: Ensures cumulative grammar progression across a course.
News-to-Lesson Pipeline. For the Skills Lab track, a Python engine ingests RSS feeds from sources like the BBC and uses Gemini 2.0 Flash to transform articles into full 10-section lessons. The pipeline enforces "Grammar Diversity" to ensure students see a wide range of linguistic structures.
PWA & Offline Support. Using Serwist, the LMS is fully installable on mobile devices with offline fallbacks for the dashboard and lesson previews, critical for professionals studying during commutes.
Credit System. AI feedback is gated on credits, which the admin can grant directly from a student's profile. This keeps compute costs predictable while allowing the teacher to comp access for students who need it.
The Curriculum
Content follows a CEFR-Task Hybrid Model. Beginner levels (A1) employ an Errorless Teaching framework (Prompt-Echo-Distractor-Transfer) to build confidence. A2 adds Proactive Exposure — vocabulary and grammar are chunked and surfaced in context before students are expected to produce them.
The curriculum currently features:
- 31-Lesson A1 Foundation: A complete sequence from Pre-A1 to A1 functional independence.
- 21-Lesson A2 Course: Business English A2 (GSE 30–36), fully interactive with chunking-zone structure and multi-mode practice blocks.
- B1 + B2 Executive Content: 5 lessons each, extending the Executive Track for higher-level professionals.
- Skills Lab (B1–C1): 15 news-driven lessons (5 per level) focusing on high-level professional discussion.
- Stealth Grammar: Grammar is never taught in isolation; it's embedded in professional phrases and supported by L1 transfer notes (cognate awareness, common interference patterns).
What I Built
- Lesson Engine: Custom renderer for 8+ interactive block types with Zod validation.
- A1 Course: 31-lesson sequence generated and verified via the Level Calibrator.
- A2 Course: 21-lesson sequence with chunking strategy, generated via a dedicated A2 eval pipeline.
- Skills Lab: End-to-end automated pipeline (Ingestion → Generation → Deployment).
- Admin Command Center: Student CRM, submission inbox, credit management, and AI-assisted feedback loop.
- AI Feedback Engine: Claude-powered drafting system with level-aware prompting and native-language (Turkish) translation appended to every response.
- Student Experience: Onboarding flow, lesson completion feedback (3-face rating), referral nudge, and in-app notification system.
- PWA Integration: Full mobile-first experience with manifest and service worker.
- Eval Harness: Automated 4-tier testing suite with level-specific rubrics (7-category for A1, 8-category for A2).
- i18n: Full English and Turkish localization via
next-intl.