The proprietary intelligence layer behind DolutechAI security products — specialized for threats with 22,000+ curated examples, SFT, high-quality DPO, and continuous RL.
SOC Model V1 is the proprietary engine inside DolutechAI security platforms. Customers access its capabilities through SOC AI Agent and SOC WebApps — not through a standalone model API.
Enterprise autonomous SOC: orchestrator, eight specialist agents, SOAR playbooks, and incident response — all powered by SOC Model V1.
Explore SOC AI Agent →Website and WebApp protection with the same model intelligence, web-only connectors, and site-friendly pricing.
Explore SOC WebApps →We do not offer public API access to the model. Intelligence is delivered exclusively through our products.
Open-weight foundation specialized into a proprietary cybersecurity model.
Curated cybersecurity examples — CVEs, incidents, playbooks, and IOC data.
Cybersecurity specialization via SFT — not a generalist with a wrapper.
High-quality Direct Preference Optimization with analyst-validated pairs.
Continuous reinforcement learning from production agentic feedback.
Human-in-the-loop active learning with governed refinement cycles.
A proven open-weight foundation paired with 22,000+ specialist-curated examples — CVE databases, incident reports, SOC playbooks, and IOC intelligence.
Domain specialization through SFT on curated cybersecurity datasets. A proprietary model trained for security — not a generalist with a security prompt.
Direct Preference Optimization with analyst-grade preference pairs aligned to how security professionals triage, investigate, and respond.
Closed-loop reinforcement learning from our production agentic pipeline — the model improves with every real-world interaction.
Decision workflow, safety controls, and the supervised learning flywheel that keeps SOC Model V1 improving in production.
Our architecture paper documents the six-stage decision workflow, deterministic screening, fail-safe escalation, and how analyst validation feeds governed model refinement.
Continuous improvement through human-in-the-loop active learning — a supervised data flywheel guided by analyst ground truth.
The model targets uncertainty and recurring noise patterns — adapting as threats evolve.
Analyst validation converts operational judgment into ground-truth labels for SFT, DPO, and governed refinement.
Deterministic screening first, fail-safe escalation on high-impact decisions, and full audit trails.
Each validated incident sharpens triage for the next — reducing alert fatigue over time.
Production event → AI recommendation → human validation → ground-truth label → curated dataset → supervised refinement → improved inference
Evaluated through our production agentic workflow — the same pipeline that powers SOC AI Agent in the field.
Internal benchmark — methodology aligned with our production agentic pipeline. Results reflect real SOC analysis tasks, not synthetic trivia.
SOC Model V1 is not a standalone chatbot — it is the proprietary intelligence layer that drives orchestration, specialist analysis, and automated response across our platforms.
Training data is specialist-curated from public CVE databases, anonymized incident patterns, and internally validated security datasets. Customer production data is not used for model training without explicit consent.
Analyst validation produces supervised ground-truth labels that feed governed refinement cycles — not risky trial-and-error on live threat decisions.
Deterministic screening, fail-safe escalation, and complete audit trails for compliance review and operational transparency.
The model powers autonomous detection, analysis, and response inside SOC AI Agent and SOC WebApps.
