JobXentral

JobXentral turns a job search into structured, per-role journeys: tailor your resume, draft the cover letter, check fit and gaps, then rehearse the interview by voice. Around them sit an ops back-office and an AI-trainer workspace for the agents behind the product.

Job seekers juggle dozens of applications with generic resumes that are not tailored to each posting or to ATS keyword scanning. Interview answers often lack structure and evidence, and LinkedIn presence stays inconsistent. On the operator side, running a multi-region product with AI agents behind it needs ops tooling and agent evaluation from day one, not as an afterthought.

JobXentral makes every application a 5-step journey: review the job description (pasted, uploaded or fetched from a URL), accept or reject proposed resume bullets, write a cover letter in a chosen tone, review strengths and gaps, then submit. Interview practice runs in Learning, Q&A and timed Trial modes. Answers are scored on structure, specificity, clarity and confidence, and a browser voice interviewer is available. A streak-based daily dashboard, resume ATS analysis and a social planner keep momentum. A region-scoped admin back-office and a 'Gym' workspace for agent versions, evals and LangGraph traces complete the platform.

• Apply Pilot journeys

  Implemented

  Each application is a persisted 5-step wizard: JD review, resume tailoring, cover letter, fit & gaps, submit. Drafts survive reloads.

• JD import from URL

  Implemented

  A server route fetches public job postings and converts them to clean text. It blocks private and loopback hosts and enforces size and time limits.

• Resume PDF parsing

  Implemented

  Uploaded resumes are parsed server-side into text for analysis, with clear errors for image-only PDFs.

• Interview prep with rubric scoring

  Partial

  Learning, Q&A and timed Trial modes grade answers on structure, specificity, clarity and confidence, then suggest a rewrite. The grader is a deterministic heuristic standing in for an LLM.

• Voice mock interview

  Implemented

  The interviewer speaks the questions and listens to the answers through the browser's speech synthesis and recognition APIs.

• Resume Engineer

  Partial

  Shows an animated ATS score ring, six category scores and keyword-gap insights next to a resume builder. The analysis currently uses mock data.

• Daily dashboard and streaks

  Implemented

  A home view combining journeys in motion, a 7-day streak, a badge ladder and the weekly LinkedIn plan.

• Admin back-office

  Implemented

  21 admin routes cover KPIs, users, CRM, campaigns, plans, audit, feature flags and CSV reports. Persona-based RBAC scopes every page to a region.

• Gym AI-trainer workspace

  Partial

  Agent versions, eval suites, run logs and a multimodal sandbox with LangGraph trace and tool-call inspection. These screens run on mock data.

• Mobile PWA

  Partial

  A mobile-first, installable port of the seeker portal with a service-worker app shell and an offline page.

• LLM chain backend

  Planned

  Eleven specified chains (extraction, role matching, resume generation, cover letters, voice rubric and more) to run behind a Python FastAPI model gateway.

Today JobXentral is a Next.js 16 / React 19 frontend prototype with three portals: Seeker, Admin and Gym. It runs on typed mock data and localStorage, and two real server routes handle PDF parsing and JD fetching. A separate Next.js PWA serves mobile. The planned production backend is fully specified: Python FastAPI services behind a model gateway, PostgreSQL with region-level row security across four schemas, Redis, and pgvector for matching and RAG.

frontend

Next.js 16React 19React CompilerTypeScriptTailwind CSS v4lucide-react

backend

Next.js Route Handlerspdf-parseFastAPI (planned)SQLAlchemy / Alembic (planned)ARQ (planned)

data

localStorage (prototype)PostgreSQL 16 with RLS (planned)pgvector (planned)Redis (planned)

ai

Web Speech API (TTS/STT)Heuristic answer graderLLM chain catalog (planned)LangGraph agent tracing (Gym UI)

infra

Docker (multi-stage, non-root)GitHub ActionsPWA service workerHetzner + Coolify (planned)

integrations

Public job posting URLsCurated job-board directoryOAuth providers (planned)Paddle (planned)

• Interview answer coaching

  Answers are scored 1 to 5 for structure (STAR), specificity (metrics), clarity and confidence (hedging detection). The feature returns strengths, improvements and a suggested rewrite, and aggregates trials into a weighted 0–100 score.

• Voice interviewer

  A speak-listen-think loop built on browser speech synthesis and recognition lets candidates rehearse out loud.

• Agent operations Gym

  Models AI agents with versions, few-shot examples, MCP/API/RAG tool bindings, canary/staging status, eval suites and LangGraph traces for multiple model providers.

• Planned LLM chains

  Eleven specified chains: resume extraction, discovery interview, top-5 role matching, tailored resumes, JD fit analysis, cover letters, interview FAQ, voice rubric and social content.

1. 01

   Safe server-side URL fetching

   The JD importer fetches arbitrary user URLs. It rejects non-HTTP schemes and private, loopback and link-local hosts, caps responses at 2 MB, aborts after 12 s and strips scripts and styles before returning text.

2. 02

   Region isolation designed from day one

   Every admin page filters by an effective region derived from a persona-based RBAC context. This mirrors the planned PostgreSQL Row-Level Security that confines regional admins to their own data.

3. 03

   Deterministic mock data as an API contract

   A seeded PRNG generates stable admin datasets, and typed mock modules define the exact shapes the future backend must return. That allows a clean frontend-first cutover.

4. 04

   Cross-browser voice capture

   The voice hook feature-detects standard and WebKit-prefixed speech recognition and manages the speaking, listening and thinking phases with graceful fallback to typed answers.

• Journey-first application model

  Applications are not rows in a tracker. Each one is a guided, resumable workflow that ends with ready-to-send materials.

• Built-in AI-ops workspace

  A dedicated Gym portal for agent versions, evals and trace inspection treats AI quality as a product surface.

• Three portals, one schema

  The Seeker, Admin and Gym portals are specified together, down to a 125-table PostgreSQL schema tracing every UI field.

• EU-first and privacy-aware

  Plans target EU hosting and EU AI providers, with GDPR and AI-Act checklists and privacy rules embedded in prompts.

The source repository is private. A live walkthrough is available on request.