config.py centralizes the hardcoded workspace path, threshold values, Ollama configuration, and filesystem locations.
main.py is the command-line entry point. It creates the logger, verifies the environment, and launches the pipeline.
parsers/pdf_parser.py extracts text from PDF files with PyMuPDF.
parsers/link_extractor.py detects URLs inside the resume text.
scrapers/common.py now owns retry-aware fetch logic, internet detection, and availability-aware scrape results.
scrapers/github.py, scrapers/linkedin.py, and scrapers/portfolio.py fetch public pages referenced by the resume and compress them into short context strings. When a source cannot be reached, the scraper returns an unavailable result instead of throwing.
agents/evaluator.py builds the candidate profile, performs deterministic feature extraction, and retries LLM evaluation before failing the resume fatally. Source failures are converted into Unavailable markers so the model still evaluates the candidate.
llm/prompts.py contains the only prompt text used by the model. The prompt demands JSON-only output.
llm/ollama.py communicates with the local Ollama HTTP API and verifies that qwen3:8b is available before processing.
exporters/excel.py writes a recruiter-friendly workbook with a summary sheet and candidate details.
core/pipeline.py orchestrates the processing flow, keeps the model calls sequential, parallelizes preprocessing, logs every step, and moves PDFs into the correct destination folder. Fatal resume failures are routed into workspace/Failed while scraper failures only emit warnings and continue.
core/queue.py provides a simple thread-safe queue abstraction for watch mode.
models/candidate.py defines the Pydantic models used across the system.
utils/logger.py configures file and console logging with timestamps.
External website failures are treated as expected runtime conditions.
The scraping layer retries each network request up to 3 times with backoff delays of 0, 1, and 2 seconds.
When a site is unavailable, the console receives a concise warning and the log file receives the full exception trace.
If internet connectivity is unavailable, Atlas skips all external scraping for the run and continues with resume-only evaluation.
Only fatal resume failures move a PDF into workspace/Failed:
- Corrupted or unreadable PDFs.
- Repeated LLM evaluation failure.
- Repeated JSON validation failure.
- Unexpected fatal pipeline errors.
flowchart LR
A[Incoming PDFs] --> B[PDF Parsing]
B --> C[Text Extraction]
C --> D[Email / Phone / Link Extraction]
D --> E[Website Scraping]
E --> F[Candidate Profile Builder]
F --> G[Deterministic Analysis]
G --> H[Qwen3 Evaluation via Ollama]
H --> I[JSON Validation]
I --> J[Excel Generation]
I --> K[Move PDF]
J --> L[Reports]
K --> M[Finished]
classDiagram
class AtlasPipeline {
+run_once() RunSummary
-_process_pdf(Path)
-_to_record(EvaluationOutcome)
-_move_pdf(Path, Path)
}
class CandidateEvaluator {
+build_profile(Path) CandidateProfile
+evaluate(CandidateProfile) CandidateEvaluation
+evaluate_path(Path) EvaluationOutcome
}
class OllamaClient {
+ensure_ready()
+evaluate_json(str) dict
+health_check() OllamaHealth
}
class ExcelExporter {
+export(list~ScreeningRecord~) Path
}
AtlasPipeline --> CandidateEvaluator
AtlasPipeline --> OllamaClient
AtlasPipeline --> ExcelExporter
- Discover PDFs in
workspace/Incoming. - Extract text and links from each file.
- Scrape public URLs when they are present, retrying failed requests lightly.
- Build a compact candidate profile.
- Apply deterministic feature extraction.
- Send one profile at a time to the local Ollama model.
- Retry and validate the returned JSON against the strict Pydantic model.
- Export the workbook.
- Move the PDF into
Shortlisted,Maybe,Rejected, orFailed.