08. Roadmap
This chapter describes capabilities that are in active development or planned for the next 12 months. The roadmap is intentionally focused on extensions of the existing architecture rather than new architectural directions; the platform's core retrieval + panel + state machine design is sufficient to support every item below.
In active development
Voice-AI candidate screening
Automated 15-minute initial screens conducted by a voice agent, with the same five-persona evaluation framework applied to the screening transcript. The output integrates directly into the candidate brief as an additional signal source, weighted alongside the resume-based panel.
The technical integration point is straightforward: the voice agent produces a transcript, the transcript is fed into the existing panel as an additional context block, and the panel rescores. No changes to the matching engine, retrieval, or pipeline state machine are required.
Predictive offer-acceptance modeling
A regression model trained on the outcome ledger that predicts the probability a candidate will accept an offer at a given comp band, given their stage preference, location, current role, and submission context. This addresses one of the most expensive failure modes in recruiting: a long process that ends with the candidate declining the offer.
Training data: every interest_confirmed → offer → hired/rejected chain in pipeline_stage_history, joined with the comp negotiations captured in recruiter_notes (note_type = 'salary').
Automated reference checks
Lightweight reference protocol where the candidate provides 2-3 references and the platform reaches out via email with structured questions. References' free-text responses are parsed into the same logo-blind quality dimensions used by the Future Peer persona, and the aggregate is added to the brief.
Agentic outreach (constrained)
The current waterfall is rule-based. The next iteration adds an agentic layer that reads each company's recent news, fundraising activity, and team page before drafting outreach, producing more contextually relevant first messages. The agent is constrained: it can only operate on already-ranked targets from the deterministic waterfall, and its output passes through the existing voice consistency engine. This preserves the cost gradient even as agentic capabilities expand.
Planned within 12 months
Multi-region deployment
Refery currently operates a single Postgres instance. As the platform extends to European clients, multi-region deployment with read replicas in EU and US regions becomes necessary. Schema-level partitioning by client region is the planned approach.
Structured brief storage
The candidate brief currently lives in candidates.ai_analysis as a text blob with consistent internal structure. Migrating the structured fields (bracket, stage fit matrix, panel verdicts, screening questions) to dedicated columns enables stronger analytics and a UI that does not need to parse text. This is a backwards-compatible migration that runs alongside the existing field.
Public API for scouts
The current scout submission flow is operator-mediated. A public API endpoint that scouts can call directly (with rate limits, signature auth, and submission validation against the structured contract) reduces operator load and accelerates the data flywheel.
Open-source signal engine components
The deterministic signal engine (logo tier classifier, trajectory analyzer, non-tech flag) does not depend on Refery's proprietary data. Releasing these as open-source modules establishes Refery as a reference implementation in the AI recruiting space and creates a recruiting effect for engineers and contributors.
The matching engine, panel system, and state machine remain proprietary.
Research directions
Calibrated bracket thresholds via outcome backtesting
The current bracket boundaries (Top 1%, Top 5%, etc.) are heuristic. Once the outcome corpus reaches sufficient size, the bracket boundaries can be calibrated against actual hire-and-retain rates. A "Top 5%" rating should correspond to an empirically measurable conversion-to-hire rate with confidence intervals. This is the formal definition of a well-calibrated rating system.
Persona prompt evolution
The five-persona prompts are currently hand-engineered. A rigorous evaluation framework, built on a held-out outcome set, would allow systematic prompt iteration with measurable quality improvement. This is the recruiting equivalent of evals-driven LLM development that AI labs use internally.
Cross-candidate pattern detection in batch runs
When the operator runs a batch of 10+ candidates simultaneously, the system surfaces patterns that single-candidate runs miss (e.g., "all 11 candidates flag stage mismatch toward Series A"). Formalizing this into a dedicated batch-mode analysis layer would surface market trends in the candidate population.
What is intentionally not on the roadmap
A few things that frequently appear on AI recruiting roadmaps and that Refery has deliberately chosen not to pursue:
- AI-generated candidate sourcing from public profiles at scale. This conflicts with consent norms and privacy expectations. Refery's growth path is the scout network and inbound, not scraping.
- Fully autonomous candidate-to-client communication. The platform drafts messages; the operator sends them. The accountability boundary is preserved.
- A general-purpose ATS. Refery is a matching and orchestration engine for a specific market segment (senior tech hiring at venture-backed startups). Becoming a general ATS would dilute the focus and degrade quality.
The roadmap is shaped by the same principle that shaped the current architecture: do the hard parts well, do not do the parts that are easy and add little.