Talent Radar

Date: April 27, 2025

Technology Stack:

Python Langchain LLM FastAPI

A self-reliant tool for navigating modern recruitment.

Version: 1.0
Date: April 2025

1. Introduction

1.1 Purpose

Hiring is repetitive. Talent Radar handles the monotony. It reads resumes, learns from job descriptions, matches intent, and prepares interviews. It works quietly and efficiently—designed to shrink time and surface relevance.

1.2 Scope

Included:

Parsing resumes, extracting signal from noise.
Matching human potential to job descriptions using vector mathematics.
Generating questions for interviews.
Updating real-time status using Redis.
Connecting with ecosystems like Odoo, WhatsApp, and email.

Not Included:

Payroll.
Onboarding.
Multi-user dashboards.

1.3 Stack

Component	Technology
Backend	FastAPI (AI), Django (Admin)
Storage	PostgreSQL
Vector Storage	Pinecone, Milvus DB
Queue	Redis
Intelligence	TogetherAI, Sentence Transformer Embeddings
Integrations	Odoo CRM

2. Shape of the System

2.1 Overview

    Resume Drop → FastAPI → Pinecone/Milvus (Vectors)
                         ↘ PostgreSQL (Metadata)
                         ↘ Redis Queue → Django Admin → Odoo / Email

2.2 Breakdown

Resume Ingestion: Accepts files, extracts text, generates embeddings.
Matching Engine: Compares job and resume vectors, ranks by closeness.
Queue Manager: Handles retries, tracks state, avoids duplication.
Admin Interface: Visual CRUD, sends data to external systems.

3. Functionality

3.1 Resume Processing

ID	Requirement	Details
FR-01	Parse files	PyPDF2, docx2txt
FR-02	Store embeddings	Pinecone + PostgreSQL
FR-03	Batch mode (100+)	Async queue

3.2 Matching

ID	Requirement	Details
FR-04	JD Embeddings	Sentence Transformer
FR-05	Rank relevance	Pincone cosine > 0.85
FR-06	Export top picks	Odoo PATCH

3.3 Interview Prep

ID	Requirement	Details
FR-07	Generate questions	Prompted metaai/llama-70b-turbo
FR-08	Deliver via Email	IMTP

4. Constraints

Category	Requirement	Implementation
Performance	50 resumes/min	FastAPI + Redis
Security	Encrypt PII	AES-256 + Django Fernet
Scaling	10K resumes/day	Kubernetes
Usability	Training < 30 min	Minimal Admin Panel

5. Endpoints

    POST /api/v1/pipeline/resume-store
    POST /api/v1/pipeline/resume-analyzer
    GET /api/v1/webhook/receiver?task_id=<ID>

    POST /api/candidates/
    GET /api/job-roles/<id>/analyze

6. Data Shapes

SQL Models

    Candidate:
      resume: file
      raw_text: text
      metadata: JSON
    
    JobRole:
      description: text
      embeddings: binary

Vector Store (Qdrant)

    {
      collection: "resumes",
      vector: [1536 floats],
      payload: {
        candidate_id: 43,
        skills: ["python", "ml"]
      }
    }

7. Failure Plans

Failure	Resolution
Parser crash	Retry → DLQ
Agent limits	Backoff 10→60s
Odoo down	Cache, then sync

8. Deployment

    docker-compose.yml:
      fastapi
      django
      qdrant

CI/CD: GitHub Actions → AWS ECS

9. Observability

Latency: P99 < 5s
Accuracy: A/B test vs HR manual
Tools: Prometheus + Grafana, Sentry

Talent Radar observes, learns, and then steps back.