AI Recruitment Platform

Overview

A full-stack AI recruitment platform that ingests résumés (PDF, DOCX, legacy DOC) and uses an LLM pipeline to extract structured candidate data into a searchable, review-ready database. It runs on a Next.js front end, a .NET and Dapper API over SQL Server stored procedures, and real-time and batch OpenAI ingestion paths.

Challenge

1. 01Parse a ~1.16M-document backlog of inconsistent CVs (PDF / DOCX / legacy DOC) into reliable records without runaway cost.
2. 02Tolerate unpredictable LLM output, from varying field names to type mismatches and missing fields.
3. 03Support both instant single-CV parsing and high-throughput, low-cost bulk processing, with dedup and cost accountability.

Approach

1. 01Built a .NET and Dapper API over 34 SQL Server stored procedures, with an ingestion worker, SHA-256 dedup, and PDF/DOCX/DOC text extraction.
2. 02Engineered a flexible LLM parser plus a real-time path (background service and channel) and a Batch-API path (submit, poll, hydrate) at ~50% lower cost.
3. 03Delivered the Next.js UI: dashboard, upload with real-time/batch toggle, candidate search, and review queue, plus the containerised CI/CD deployment.

Outcomes

1. 01Verified end-to-end across both real-time and batch ingestion paths.
2. 02Held blended cost near ~$0.006 per CV at scale, with ~79% of the 1.16M-document backlog processed.
3. 03Built-in cost observability and resilient field mapping that recovered skills and location the LLM emitted under inconsistent keys.