CandiSift

CandiSift is an open-source applicant tracking system for teams screening resumes with an AI-assisted workflow. It is best suited to recruiters prepared to self-host and configure the system; its structured review and audit trail are useful, but its rankings come without a shipped accuracy benchmark.

Overview

CandiSift takes resumes through parsing and PII removal, applies hard filters, ranks candidates, and sends those who remain to persona agents and recruiter review. Recruiters can inspect ranked results and accept or reject candidates rather than relying on an unattended screening decision. That makes the system a stronger fit for teams that want a defined review process than for those seeking validated automated selection.

It accepts PDF, DOCX, and plain-text resumes, including multi-file uploads. Ranking uses token-cosine similarity, with technical and risk persona agents and synthesis supporting the review. Claude integration runs through Agno persona agents; an offline stub LLM can be used without an API key, though that does not establish ranking quality.

Key features

• Filters before model processing: Deterministic hard filters run before the paid LLM stage. This can reduce unnecessary model processing, while keeping initial screening rules distinct from later AI analysis.
• Staged batch cost estimates: Users can see an estimated processing cost before confirming a batch. That offers a useful checkpoint for teams managing variable per-token charges, though the estimate does not remove those charges.
• Recruiter decisions and audit: An append-only log records scores, rationales, decisions, and versions. Along with ranked results and accept-or-reject controls, this supports review of how a candidate moved through the workflow.
• Privacy and bias controls: CandiSift says it removes personally identifiable information before model processing and provides a bias-audit endpoint covering pass-through and recommendation mix. These controls are relevant to teams scrutinizing screening decisions, but they do not substitute for an accuracy benchmark.
• Security measures: The project describes HTTP Basic authentication, security headers, rate limiting, request-body limits, upload validation, and prompt-injection defenses. These are practical safeguards for an application handling candidate documents; self-hosting teams still need to manage deployment and configuration.
• Deployment and storage: It can run in Docker, stores state in SQLite by default, and supports a configurable Postgres database URL. The project has no shipped accuracy benchmark, and OCR, embeddings/FAISS, and skillNER are documented swaps rather than default capabilities, so teams needing those functions must account for their absence.

Pricing

Open-source self-hosted: free. There is no software fee, but LLM processing can incur per-token costs. This plan suits teams able to host and configure the application and manage model usage. It is not a hosted subscription with those operational responsibilities bundled in.

The cost estimate before confirming staged batches and the hard-filter-first workflow give users some control over paid LLM processing. Teams that need OCR, embeddings/FAISS, or skillNER by default should look elsewhere or plan for those capabilities separately.

Platforms

CandiSift supports API, self-hosted, and web access. Docker deployment and SQLite by default, with Postgres configurable, give teams a choice of deployment and storage setup. This is a practical fit for organizations comfortable operating their own software, rather than those looking for a managed recruiting service.

Who it's for

The intended workflow is for recruiters creating roles, uploading resumes, reviewing ranked candidates, and accepting or rejecting them. It suits teams that value control over screening stages, cost checkpoints, and an audit record, and can take responsibility for self-hosting. It is a weaker fit for teams that require established ranking-accuracy evidence or expect OCR and other documented swaps to be ready to use by default.

Pros and cons

• Pro: Hard filters precede the paid LLM stage, helping limit unnecessary model processing.
• Pro: Recruiter decisions and an append-only log make screening outcomes and their recorded rationales reviewable.
• Pro: PII removal and a bias-audit endpoint provide specific privacy and screening-review mechanisms.
• Con: There is no shipped accuracy benchmark, so teams cannot use a published measure of ranking performance to judge its suitability.
• Con: OCR, embeddings/FAISS, and skillNER are not included default capabilities, limiting out-of-the-box coverage for teams that need them.
• Con: LLM processing can add per-token costs, and the self-hosted setup puts deployment responsibility on the team.

Alternatives

For a broader comparison, browse AI Recruiting Assistants or Candidate Screening Software.

• Metaview is a better candidate for teams seeking sourcing profiles and concurrent searches, with Pro at $100 a month billed monthly per user and 200 profiles sourced per month.
• Sprad suits teams wanting a free ATS foundation: ATS Core is free, billed €0 forever, with career-site, pipeline, team, analytics, and MCP/AI connection features.
• Avery is worth considering for European profile discovery, offering ranked matches and market views; its free plan excludes outreach and contact details.
• HireMIQ may suit teams looking for a broader AI recruiting toolkit, with a free plan that includes eight AI modules, Boolean search, candidate scoring, outreach templates, and assessments.
• Pin is an alternative for contact lookup and email sequences, though its free plan is capped at five lookup credits per week, one job position, and one AI agent.
• Hirin is another recruiting platform to compare.
• Talview may fit organizations seeking unit-based tailored pricing, with a typical minimum commitment of approximately 150 to 200 candidates.
• HireXAI is another paid web-based alternative.

Verdict

CandiSift is a sensible choice for recruiters who can self-host and want a controllable screening workflow with cost checkpoints, human decisions, and an audit trail. Its main reason to choose it is that combination of configurable processing and review; its main reason to look elsewhere is the absence of a shipped accuracy benchmark, especially if ranking quality must be demonstrated before adoption.