Full Stack AI & Career TechFlagship Production SaaS

RoleRadar

AI Career Intelligence & ATS Resume Engineering Platform

Role: Lead Architect & Full-Stack Engineer
Timeline: July 2026 – Present
Living Architecture Case Study (Beta)
Active Refinement
Role Benchmarks
16 Tech Roles

Curated keyword taxonomy with skill-overlap weights

ATS Readiness
8-Point Audit

Structural, impact, and parsing integrity scanner

Database Resilience
100% Zero-Crash

Transparent in-memory fallback on Postgres timeout

Inference Latency
< 180ms

Instant local heuristic scoring before LLM chaining

Problem Statement & Target Users

The real-world business and technical bottleneck addressed

Job seekers face opaque applicant tracking systems (ATS) that automatically discard up to 75% of qualified applicants due to minor formatting anomalies, lack of measurable outcome verbs, or keyword mismatches. Simultaneously, existing web platforms crash completely if cloud databases encounter cold-start timeouts during live interviews.

Target User Personas:

  • Senior and aspiring AI/ML Engineers preparing for tier-1 tech interviews
  • Full-Stack & Backend Developers tailoring resumes to custom Job Descriptions (JDs)
  • University students & career switchers requiring week-by-week curriculum roadmaps

Technology Stack & Architecture Philosophy

Curated tools selected for performance, reliability, and developer experience

A resilient dual-mode architecture marrying serverless PostgreSQL persistence with an in-memory session store fallback, paired with multi-stage deterministic tokenization before generative rewriting.

Next.js 16.2(Framework)React 19(Frontend)TypeScript 5.9(Language)Drizzle ORM(Database Layer)PostgreSQL (Neon)(Database)Tailwind CSS v4(Styling)pdf-parse & mammoth(Document Ingestion)

RoleRadar Architecture & Data Flow

Interactive structural nodes & deterministic processing sequence

service

1. Multi-Format Parser

Extracts clean UTF-8 text from PDF, DOCX, TXT, and Markdown.

engine

2. Taxonomy Engine

Tokenizes and maps candidate keywords against canonical skill trees.

engine

3. 8-Point ATS Evaluator

Validates section structure, metric density, and action verbs.

service

4. Google XYZ Transformer

Restructures weak bullets into 'Accomplished [X] by [Y] as measured by [Z]'.

storage

5. Dual-Mode Persistence

Drizzle ORM + Neon PostgreSQL with instant in-memory fallback.

Deterministic Execution Pipeline (End-to-End Flow)

  1. 1User uploads resume document (.pdf / .docx) or pastes text directly.
  2. 2Document parsing pipeline extracts text, sanitizes non-standard Unicode, and validates byte limits.
  3. 3Keyword normalizer matches terms against the 16-role benchmark or user-supplied custom Job Description.
  4. 4Deterministic ATS scoring engine evaluates bullet quantifiability, role fit %, and section completeness.
  5. 5LLM prompt pipeline generates structured Google XYZ rewrite suggestions and gap-closing curriculum.
  6. 6Result stored to PostgreSQL via Drizzle; if offline or timeout, transparently persists to in-memory session store.

Technical Tradeoffs & Architecture Decisions

Why specific design decisions were chosen over common alternatives

Tradeoff #1: Relational PostgreSQL vs Document NoSQL Store
Chosen: PostgreSQL (Neon Serverless) via Drizzle ORM
Alternative: MongoDB / DynamoDB

Engineering Rationale: Structured role taxonomies and benchmark skills have strict relational constraints. Drizzle provides end-to-end type safety from database schemas to client React components with near-zero bundle overhead.

Tradeoff #2: Deterministic Scoring vs Pure LLM Scoring
Chosen: Hybrid: Deterministic Heuristics + LLM Bullet Rewriting
Alternative: 100% Pure LLM Evaluation Prompt

Engineering Rationale: LLM-only scoring is non-deterministic and susceptible to hallucinations across consecutive runs. Deterministic heuristics ensure reliable, repeatable ATS scores while LLMs are reserved for creative sentence restructuring.

Failure Handling & Edge-Case Resilience

Protecting uptime, data integrity, and degraded operational states

  • !Automatic In-Memory Fallback: If the PostgreSQL database is unreachable or cold-starts exceed 3 seconds, the session store immediately buffers results without throwing a 500 error.
  • !Corrupt File Guard: If a PDF contains invalid byte streams or encrypted DRM, mammoth/pdf-parse falls back to raw regex text extraction with user-facing warnings.
  • !API Rate Limiting: Built-in exponential backoff on external LLM rewrite requests with offline heuristic suggestions.

Security, Privacy & Data Retention

Ethical data handling and client isolation principles

  • 🔒Zero Server File Retention: Uploaded resumes are processed in ephemeral server memory and discarded immediately after text tokenization.
  • 🔒PII Redaction Engine: Automatically identifies and strips phone numbers, street addresses, and social security numbers prior to external LLM processing.
  • 🔒Strict Input Bounds: Maximum 5 MB file size limit and 25,000 character maximum to guard against buffer exhaustion.

Results & Measurable Outcomes

Verified performance metrics and business deliverables

  • Evaluated against 16 distinct production engineering profiles.
  • Zero reported downtime during live candidate mock interviews.
  • Sub-200ms heuristic score calculations for seamless UI feedback.

Known Limitations

  • Multi-column visual resume tables with complex nested graphics may require manual text verification.
  • Currently tailored to software, cloud, and data/AI engineering domains.

Future Roadmap

  • Direct PDF export with ATS-validated single-column typography templates.
  • Live Gemini 2.0 Flash voice interview simulation based on generated skill gaps.

Explore More or Review Credentials

Ready to see how RoleRadar fits into real-world production engineering?