An autonomous multi-agent system implementing the full hiring cycle β from requirements analysis to final candidate evaluation β without human intervention.
Four Scientific Foundations
The platform is built on four key scientific-technical components that make autonomous, evidential hiring possible.
CALM Semantic Scoring
Context-aware latent matching β finding candidates by the meaning of skills and experience, not keyword overlap.
A2A / MCP Architecture
Decentralized multi-agent architecture where specialized agents collaborate autonomously: Analyst, Selector, Interviewer, Evaluator.
GRPO Strategy Optimization
Group Relative Policy Optimization (arXiv:2402.03300) trains interview agents to evolve multi-step dialogue strategies, not just record responses.
ICONIC I-Stack v2.0
Six-vector evidential assessment model that replaces resume self-reporting with verified artifacts of real professional activity.
GRPO: Self-Learning Interview Agents
Existing HR systems have static logic, incapable of adapting to changing market requirements. Iconicompany introduces GRPO (Group Relative Policy Optimization) β a breakthrough from arXiv:2402.03300 β to train autonomous multi-step interview strategies.
Dialogue Strategy Optimization
Unlike systems that merely record and format static responses, GRPO transitions from speech recording to optimization of multi-step dialogue trajectories. The agent generates groups of different interview scenarios, compares their effectiveness, and selects question chains that most accurately verify candidate skills.
Outcome-Based Learning
GRPO teaches the agent to evolve: it discovers unique ways to detect candidate "hallucinations" or AI-assisted cheating in real time. The system learns from real outcomes β whether a candidate passed their probationary period β rather than expensive manual annotation.
30β40% Compute Cost Reduction
Applying the mathematics from arXiv:2402.03300 eliminates the need for a classical Reward Model, which is critical for reducing computational complexity and removing the costly manual annotation step. This results in 30β40% lower fine-tuning compute costs compared to foreign analogs.
No Reward Model Required
The GRPO contour optimizes agent dialogue strategies without constructing a classical Reward Model β critical for reducing computational complexity in production deployment.
Citation: DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300
ICONIC I-Stack v2.0
Classical hiring relies on resume self-presentation β an unstructured text document not amenable to automatic verification. Over 70% of resumes contain exaggerations or inaccurate skill claims (Mercor, HeroHunt). Iconicompany implements evidential hiring: principled rejection of resume text as the primary source of candidate information. Instead, a unified digital identity is formed from real artifacts of professional activity.
Architectural Impulse
Assessment of systems thinking through C4 diagram analysis (Draw.io/Mermaid). Models (Kimi 2.5) verify component connectivity, database choices, and integration patterns. Provides verification of architectural competency.
Logic & Construct
Expert assessment of the candidate's ability to critically audit a Pull Request generated by an AI agent (DSPy) with deliberate architectural and logical defects. An "Expert-as-a-Judge" model (Kimi k2) evaluates not just defect detection, but the depth of argumentation in PR comments β a precise signal of cognitive vigilance when working with potentially hallucinating LLM outputs.
Source Integrity
Verification of authentic candidate thinking: recording and semantic analysis of the reasoning chain during a live session. The system evaluates the cognitive trajectory of problem-solving β correct logic with an erroneous final result is scored as a positive signal. Creates a controlled environment for verifying genuine engineering thinking in real time.
Energy Score
Semantic alignment with the company tech stack: finding semantic connections between the candidate's stack and the company's tech debt/tasks in latent space β not simple skill list intersection. Delivers true contextual fit measurement.
Technical Intelligence
Real-time analysis of terminology density and answer depth (voice/text stream). The density and contextual accuracy of professional vocabulary serves as a proxy indicator of actual competency level β impossible to fake with AI assistance.
Capability & Growth
Growth potential and leadership competencies: assessment of Code Review skills, mentorship capabilities, and architectural thinking based on behavioral patterns in dialogue. Identifies senior readiness and trajectory.
All six vectors aggregate into a calibrated Compliance Indexβ’ (0β100) with a Brier Score β€ 0.15. Together these components form a closed optimization loop where each completed hire becomes a training signal for the system.
Autonomous Agent Pipeline
Four specialized agents work in sequence to deliver a verified team configuration within 48 hours β with Human-in-the-loop control at every stage.
Analyst Agent
Decomposes the project brief into engineering roles, latent skill vectors, and a structured requirements map in 0β10 minutes.
Requirements arrive as free-text briefs. Manual decomposition into roles and competencies takes days and introduces bias.
The agent applies CALM semantic scoring to parse the brief, identify implicit skill requirements, and produce a structured role map with latent vector profiles for each position.
Key Benefits
- Decomposition in 0β10 minutes vs days manually
- Latent skill vectors β not just keyword lists
- Human-in-the-loop review before sourcing begins
Selection Agent
Scans 1M+ profiles using Energy Score (I-Nexus) pre-screening to surface the highest-relevance candidates automatically in 10β60 minutes.
Manual market scanning across multiple platforms is slow, incomplete, and biased toward visible (not best-fit) candidates.
The Selection Agent scans aggregated databases, applies Energy Score latent-space matching, and pre-scores candidates against the role vectors generated by the Analyst Agent.
Key Benefits
- 1M+ profiles scanned in under 60 minutes
- Energy Score semantic matching β not keyword filters
- Only high-relevance candidates proceed to interview stage
Interview Agent
Conducts autonomous technical interview sessions with GRPO-optimized dialogue strategies, applying I-Origin and I-Code verification in 1β24 hours.
Human interviewers are inconsistent, expensive, and can be gamed with AI-assisted answers. At scale, manual interviews become the bottleneck.
The GRPO-trained Interview Agent runs multi-step technical sessions, records reasoning chains (I-Origin), audits PR artifacts (I-Code), and verifies terminology density (I-Interview) β detecting AI-assistance in real time.
Key Benefits
- GRPO-optimized dialogue β evolves with every hire
- AI-assistance detection in real time (I-Origin)
- Senior verifiers join only for final complex competency checks via MatrixRTC
Compliance Indexβ’ Agent
Aggregates all six I-Stack vectors into a calibrated Compliance Indexβ’ (0β100, Brier Score β€ 0.15) and generates the final team configuration report.
Final hiring decisions are made on incomplete, subjective data. Stakeholders lack a single authoritative document to compare candidates objectively.
The Compliance Agent aggregates I-Impulse, I-Code, I-Origin, I-Nexus, I-Interview, and I-Climb scores into a calibrated index, validates against Brier Score thresholds, and produces a Team Repository configuration report.
Key Benefits
- Calibrated 0β100 index with Brier Score β€ 0.15
- Full Team Repository report delivered in 24β48 hours
- Each completed hire feeds back into GRPO training loop
Product Roadmap
Stage 1: Foundation
- Analyst Agent: brief decomposition and latent role vector mapping.
- Selection Agent: Energy Score pre-screening of 1M+ profiles.
- Interview Agent (GRPO v1): I-Origin and I-Code autonomous sessions.
- Compliance Indexβ’: 6-vector aggregation and Team Repository reports.
Stage 2: Calibration
- GRPO feedback loop: probationary period outcomes β strategy retraining.
- I-Climb vector: leadership & growth assessment at scale.
- MatrixRTC Expert Pulse: senior verifier integration for complex roles.
- Compliance Indexβ’ Brier Score calibration on real hire outcomes.
Stage 3: Enterprise Scale
- Full Vendor Swap model: replace external IT integrators end-to-end.
- CALM v2: cross-company latent skill graph for market-wide benchmarking.
- Predictive I-Climb analytics: identify future tech leads before they self-identify.
- Multi-client GRPO federation: shared strategy optimization across enterprises.