ResuPulse Architecture
An AI-powered resume analysis platform combining unstructured document processing, seven deterministic scoring engines, and eight parallel LLM specialist agents into an asynchronous high-throughput pipeline.
Why Pure LLMs and Simple Regex Both Fail at Resume Analysis
The fundamental architectural dilemma between probabilistic AI hallucinations and rigid regex pattern matchers.
Hallucinations, Cost & Inconsistent Scoring
Passing an entire raw PDF resume to a single LLM prompt results in high token costs, slow latency (15-30s), and volatile scoring variance (the same resume can receive scores varying from 65 to 88 across runs). Single-prompt LLMs also fail to reliably count words, compute margin distributions, or detect precise layout geometry issues.
Zero Semantic Depth & Rigid Keyword Stuffing
Legacy ATS checkers rely solely on keyword matching and regex patterns. They cannot evaluate whether a candidate actually demonstrated engineering leadership versus simply listing a buzzword. They offer no meaningful qualitative advice, no recruiter perspective, and no grounded bullet point rewrites.
End-to-End Multi-Agent Dataflow & Queue DAG
How unstructured candidate resumes travel from raw bytes through privacy sanitization, deterministic engines, and parallel agent execution to SSE client delivery.
The Hybrid Architecture: 7 Deterministic Engines + 8 Parallel Specialist Agents
Combining mathematical certainty for objective facts with specialized LLM reasoning for qualitative diagnosis.
7 Deterministic Engines (TypeScript)
Zero Hallucinations8 Parallel Specialist Agents (Groq LLM)
Structured SchemaProduction Platform & Analysis Interface
Authentic captures of the multi-dimensional scoring report, interactive resume builder, ATS stream extractor, and job match matrix.
AI Resume Analysis & Multi-Dimensional Report
Analysis ReportComprehensive candidate evaluation with 7-dimension scoring breakdown, ATS compatibility verdict, recruiter perspective, and high-priority remediation roadmap.
Interactive Real-Time Resume Builder
Resume BuilderDecoupled 3-axis resume engine featuring 14 typography styles, 19 career categories, and live authentic A4 document rendering.
ATS Parse Diagnostics & Machine Stream
ATS DiagnosticsIn-depth ATS parsing test: unclipped spatial text stream extraction, section header detection, font subset verification, and machine readability health.
Compare Job & Requirements Evidence Matrix
Job MatchCandidate qualification comparison against target job postings, requirements evidence matching, and recruiter screening likelihood.
Detailed Pipeline Execution Stages
In-depth breakdown of how each asynchronous worker processes candidate payloads.
Document Ingestion & Geometry Analysis
Spatial Parsing & OCR FallbackDeconstructs unstructured PDF documents using bounding-box spatial geometry, text clustering, and font-weight classification. Detects multi-column rails, irregular tables, and initiates OCR fallback when vector text streams are unreadable.
Pre-Inference Structural PII Redaction
In-Memory Privacy FilterPrior to invoking external LLM providers, all candidate phone numbers, personal email addresses, home addresses, and candidate identifiers are scrubbed and anonymized in-memory. Zero sensitive personal identifiers reach inference endpoints.
Deterministic Fact Extraction
7 Deterministic Scoring EnginesRules-based TypeScript engines analyze the sanitized document corpus. They calculate factual metrics: page count, margin ratios, metric densities, active/passive verb counts, and skill taxonomy matches with zero hallucination risk.
Distributed Queuing & Concurrency Control
BullMQ Priority Queues on RedisWorkloads are enqueued into BullMQ priority lanes on AWS ElastiCache Redis. Concurrency limiters and rate-limiting token buckets prevent upstream LLM provider throttling while ensuring fair resource allocation across Free and Pro tiers.
Parallel Specialist LLM Inference
8 Specialist Agents via GroqDeterministic facts and sanitized resume sections are dispatched in parallel to 8 domain-specific LLM agents with structured JSON schema contracts. Downstream specialists reason over specific resume aspects concurrently rather than in a single bloated prompt.
Real-Time Streaming & Canonical Binder
Redis Pub/Sub & Server-Sent EventsStage progress and agent outputs are streamed to the client in real time via Server-Sent Events (SSE) backed by Redis Pub/Sub. The final orchestrator binds engine facts and specialist verdicts into a strict canonical response contract.
Production Cloud Operations
Architected for zero downtime, strict rate limit compliance, and bounded compute cost.
AWS EC2 • PM2 Cluster
Deployed on AWS EC2 (Mumbai region) running a PM2 process manager cluster with zero-downtime rolling reload. Cloudflare sits in front as an edge proxy, handling SSL, caching, and DDOS mitigation.
AWS ElastiCache Redis
ElastiCache Redis powers BullMQ asynchronous worker queues with distinct priority lanes, token-bucket rate limiters for LLM calls, and Redis Pub/Sub for SSE client progress streaming.
Format A Dual-Plane Gating
Server-side field-level access projection (`{ value: T, locked: boolean }`). Free users receive server-blanked locked fields with zero paid data transmitted over the wire, preventing client-side inspection bypass.
Explore the Complete Engineering Specifications
Comprehensive system design documentation, multi-agent sequence diagrams, distributed queue state machines, and canonical API schemas are published in the open engineering showcase repository.