EraMatch AI Features Trials

Research, experiments, and model fine-tuning for EraMatch's AI recruitment features.

Structure

1. GitHub Inspired Question

ACE-Recruiter engine: evaluating developer profiles using GitHub-inspired rubrics, parallelization, and latency optimization.

2. Question Generation and Extraction

Automated QAG and answer evaluation. Fine-tuning notebooks for general QAG, MCQ, and essay scoring (Llama, Qwen, T5, RoBERTa).

3. Job Description QAG & Keywords Transformation

JD analysis: keyword extraction and transformation into screening questions. Fine-tuning notebooks and data generation pipelines.

4. Behavioural Analysis

Behavioral trait parsing and analysis from candidate interviews.

5. CV Analysis

CV parsing, feature extraction, and JD alignment.

• dataset/ — v3 and v4 generation pipelines, generated data, EraParse 4,950-CV manifest, v4 dataset (split zip)
• benchmarking/ — Extraction + LLM structuring benchmarks, Kaggle kernel outputs
• phase_1_finetunes/ — QLoRA Gemma-3-1B fine-tune on CVSchema, benchmark results
• EraParse/ — Next-gen CV parsing pipeline (under construction)
• scripts/ — CV-JD alignment generation

6. Code Generation

Automated code generation and LeetCode dataset analysis.

7. Code Plagiarism Detection

Code plagiarism and similarity detection experiments.

8. Avatar Detection

AI-generated face detection for live interview verification.

• src/ — Dual-branch model (DCT + SRM frequency features, RGB spatial)
• notebooks/ — EfficientNet → CNN-LSTM → multimodal → DCT frequency domain
• modal/ — Serverless training (DCT, SRM, cross-modal, video-level)
• kaggle/ — SRM + ConvNeXt deployment trials
• tests/ — Pytest suite
• results/ — Thesis summary, checkpoints, plots

---

See subdirectory READMEs for details on each module.