ARIN Platform
Autonomous Risk Intelligence Network — 22 specialized AI agents form a decision council that independently analyzes every entity and produces a single explainable verdict (BUY / SELL / HOLD / AVOID) with full consensus, dissent map, and confidence score. Meta-Decision Governor, Entity Memory, Learning Agent, Product Accelerators. NVIDIA NIM inference, NeMo Guardrails, GNN systemic risk mapping.
System Overview
ARIN Platform is the multi-agent AI decision engine at the core of SAA Alliance infrastructure. Unlike single-model architectures, ARIN implements a council paradigm: 22 specialized risk agents — each with distinct analytical mandates, independent reasoning chains, and domain-specific calibration — process every entity in parallel and submit individual assessments to the Meta-Decision Governor.
The Governor aggregates agent outputs using Confidence-Reliability Weighted Mean consensus, tracks dissent patterns, and produces a single explainable verdict (BUY / SELL / HOLD / AVOID) with a composite confidence score, full dissent map, and regulatory audit trail. The Entity Memory module maintains persistent context across sessions — tracking how risk profiles evolve over time, enabling longitudinal analysis and drift detection. The Learning Agent monitors prediction accuracy and continuously recalibrates agent weights based on historical performance.
All inference runs through NVIDIA NIM API (DeepSeek R1, Llama 3.3 70B, Llama 3.1 8B/70B, Nemotron) with NeMo Guardrails for compliance, safety, and factuality filtering. GPT-4o serves as a fallback provider. Graph analysis leverages PyTorch Geometric and DGL for GNN-based systemic risk detection, dependency mapping, and cascade effect simulation across interconnected entities.
Agent Roster — 22 Specialized Units
Each agent operates independently with its own reasoning chain, domain-calibrated prompts, and reliability score. The council architecture ensures no single point of analytical failure.
Meta-Decision Architecture
Meta-Decision Governor
Aggregates 22 agent assessments using Confidence-Reliability Weighted Mean. Implements Hard Routing for domain-specific queries, Zero-Outcome Freeze for insufficient data, and Dual-Publish for contested verdicts. Produces final verdict (BUY/SELL/HOLD/AVOID) with composite confidence and full dissent map.
Entity Memory
Persistent per-entity analysis database (agent_entity_memories) tracking every risk assessment across sessions. Before each new analysis, agents receive historical context: trend direction (rising/falling/stable), anomaly detection (Z-score > 2σ flags deviations), score delta vs. previous, and 7-day trend slope. After each verdict, the analysis snapshot is saved with key factors and reasoning summary. Enables longitudinal drift detection — the system knows «entity X risk has been rising for 5 consecutive analyses» before agents even begin their assessment.
Outcome Feedback Loop
Every verdict is pre-recorded in agent_outcomes with predicted score and direction. When actual outcomes arrive via POST /api/v1/system/outcome, the system computes Brier Score per agent, triggers Platt Scaling recalibration (minimum 10 outcomes per agent), and updates the adaptive coupling matrix. The consensus engine auto-loads Platt parameters with 60-second TTL caching. Red Team adversarial validation blocks verdicts with excessive error probability. Without outcomes: frozen weights. With outcomes: self-calibrating system.
RAG Knowledge Corpus
12 structured knowledge chunks across 6 domains injected into agent context before assessment: Financial Regulation (Basel III CET1/LCR/NSFR, Solvency II SCR, TCFD/ISSB S2, IFRS 9 ECL), Climate Science (IPCC AR6 tipping points, NGFS scenarios, physical risk AAL/PML), Geopolitical (conflict escalation ladder, supply chain chokepoints), Cyber Security (NIST CSF 2.0, MITRE ATT&CK), Macro Economics (recession indicators, business cycle), Supply Chain (disruption taxonomy JIT/JIC). Domain auto-matched by object type — financial entities get Basel III, climate assessments get IPCC.
Learning Agent
Continuously monitors verdict accuracy against realized outcomes. Adjusts agent reliability weights using Platt Scaling (sigmoid calibration of raw scores) and Isotonic Regression (monotonic probability mapping). Computes Expected Calibration Error (ECE) to detect overconfidence. Red Team validation — adversarial challenge based on dissent strength, error probability, and reliability metrics. Identifies degrading agents and triggers automatic weight recalibration. 5 calibration runs for adaptive coupling matrix convergence.
Historical Events Engine
52+ curated historical events across 5 domains feeding agent stress-test context: Financial crises (Asian 1997, LTCM, Dot-com, Euro Debt, Black Monday, Barings, Tequila, Argentine, Russian), Wars (Ukraine 2022, Syria, Gulf War, Taiwan Strait), Climate (Katrina, Maria, Bushfires, Pakistan Floods, Tohoku, Heatwave 2003), Pandemics (SARS, MERS, Ebola, H1N1), Cyber attacks (SolarWinds, Colonial Pipeline, NotPetya, WannaCry, Equifax). Each event includes severity, financial loss, casualties, cascade effects, lessons learned, and recovery timeline.
Closed-Loop Agent Architecture
Every assessment flows through a 6-stage closed loop. No service is isolated — data cascades from memory to knowledge to agents to consensus to feedback and back to memory:
① Load → ② Enrich
① Entity Memory Load: Before assessment, load previous analyses for the entity — trend direction, anomaly flags, historical scores. ② Knowledge Injection: RAG corpus auto-selects relevant domain knowledge (Basel III for banks, IPCC for climate) and injects into agent context. Agents start every analysis with full institutional memory.
③ Assess → ④ Consensus
③ Agent Assessments: 22 agents run in parallel with enriched context — each seeing entity history + domain knowledge. ④ Consensus Engine: Platt-calibrated confidence-weighted aggregation with dissent detection, Red Team adversarial validation, and adaptive coupling from outcome feedback. Produces DecisionObject with full provenance chain.
⑤ Record → ⑥ Save
⑤ Outcome Pre-Record: Verdict saved to outcome table. When actual results arrive (manually or via automation), Brier Score computed → Platt recalibration triggered → coupling matrix updated. ⑥ Entity Memory Save: Current analysis persisted with delta, trend, key factors. Next assessment cycle starts at ① with accumulated intelligence.
Consensus Mechanics
ARIN does not use simple majority voting. The Governor employs a multi-layered consensus protocol designed for adversarial robustness:
- Hard Routing — Domain-specific queries are routed to relevant agent subsets. A credit analysis query activates Credit, Counterparty, Concentration, and Macro-Economic agents; irrelevant agents (e.g., Cyber, Supply Chain) are excluded from the voting pool to prevent noise dilution
- Confidence-Reliability Weighted Mean (CRWM) — Each agent’s vote is weighted by two factors: (a) its self-reported confidence for the current query, and (b) its historical reliability score maintained by the Learning Agent. High-confidence + high-reliability votes dominate the consensus
- Zero-Outcome Freeze — If no agent achieves confidence above the configurable threshold (default: 0.45), the Governor refuses to issue a verdict and flags the entity for manual review. This prevents false certainty in data-scarce scenarios
- Dual-Publish Protocol — When the council is deeply split (e.g., 11 agents BUY vs. 11 agents SELL), the Governor publishes both majority and minority verdicts with separate confidence scores, explicit reasoning chains, and a «CONTESTED» flag. No forced resolution of genuine analytical disagreement
- Dissent Mapping — Every verdict includes a full dissent map showing which agents agreed, disagreed, and abstained, with individual reasoning summaries. Enables regulatory audit and explainability requirements (EU AI Act, SEC, MiFID II)
Unified Integration Layer
ARIN is the convergence point for the entire SAA Alliance ecosystem. All seven production modules export analytical data to ARIN via the Export API, enabling cross-domain synthesis:
Inbound Data Sources
- Risk Analyzer — VaR/CVaR, stress test results, portfolio metrics, factor decomposition
- Investment Dashboard — V2 institutional reports, Report Agent outputs, sector scans, watchlist signals
- News Analytics — Sentiment scores, narrative events, RiskMirror alerts, impact graph clusters
- Crypto Analytics — Token ratings (v1.4), DeFi protocol snapshots, on-chain anomalies
- Global Risk Platform — Country/city/enterprise risk scores, 280-factor Unified Stress Report, digital twin outputs
- External Feeds — USGS, FEMA, WHO, World Bank, IMF, OFAC SDN, CISA KEV (real-time ingestion)
Output & Export
- Verdict Object — JSON: verdict, confidence, dissent_map, agent_assessments[], entity_memory_ref, audit_hash
- Export Storage — PostgreSQL persistence with entity binding (portfolio, symbol, ISIN, country code)
- API Endpoints —
POST /api/v1/unified/export,GET /api/v1/unified/verdict/{entity_id} - WebSocket Stream — Real-time verdict updates for portfolio monitoring dashboards
- Audit Trail — Immutable log of every agent assessment, Governor decision, and data input for regulatory compliance
AI Engine & Infrastructure
LLM Inference
Primary: NVIDIA NIM API — DeepSeek R1 (reasoning), Llama 3.3 70B (general analysis), Llama 3.1 8B (fast routing), Llama 3.1 70B (deep analysis), Nemotron (structured extraction). Fallback: GPT-4o. Response caching (Redis), retry with exponential backoff, multi-provider load balancing. NeMo Guardrails for compliance filtering, toxicity prevention, and factuality enforcement.
Graph & ML
Graph: Neo4j (persistent graph DB), NetworkX (analysis), PyVis (visualization), PyTorch Geometric / DGL (GNN for systemic risk and contagion). ML: XGBoost (credit scoring, classification), PyTorch (deep learning). Data: pandas, numpy, polars, scipy. All compute GPU-optimized via NVIDIA CUDA.
Backend Stack
Python 3.11+, FastAPI, Uvicorn (ASGI), Pydantic 2 (strict schema validation). SQLAlchemy 2.x + asyncpg + Alembic (migrations). PostgreSQL (users, roles, verdicts, exports). Redis (cache, Entity Memory hot store). Celery (async task queue). JWT (python-jose), Passlib (bcrypt), OAuth-ready.
Frontend Stack
Next.js 14, React 18, TypeScript. Tailwind CSS (Midnight Command design system). ECharts, D3.js, Plotly, Three.js, Recharts, react-force-graph-2d/3d (decision graph visualization). Zustand (state), TanStack React Query (data fetching), Axios. Full responsive design.
Compliance & Audit
- Authentication — Registration, login, JWT tokens, OAuth-ready; role-based access control (RBAC) — Admin, Analyst, Viewer; users and roles in PostgreSQL with Alembic migrations
- GDPR Compliance — Data access, export (JSON/CSV), deletion (right to erasure); audit log for all data processing operations; configurable retention policies
- Regulatory Audit Trail — Immutable log of every agent assessment, consensus computation, Governor decision, and data input. Designed for SEC, MiFID II, EU AI Act, and DORA compliance review
- NeMo Guardrails — NVIDIA NeMo safety framework prevents hallucination, toxicity, and non-compliant outputs. All LLM responses pass through guardrail validation before reaching the verdict pipeline
- Backup & Recovery — Automated PostgreSQL backups, Redis persistence, disaster recovery procedures; 99.8% uptime SLO target
API Reference
Base: /api/v1 · Auth: Bearer JWT · Format: JSON
- Health —
GET /health,GET /api/v1/performance/health - Agents —
GET /api/v1/agents(list all 22),GET /api/v1/agents/{id}(agent detail + reliability score) - Analysis —
POST /api/v1/risks/analyze— triggers full council assessment; params: entity_id, analysis_type, agent_subset (optional) - Graph —
GET /api/v1/graph,GET /api/v1/graph/visualization— entity dependency graph, systemic risk clusters - Auth —
POST /api/v1/auth/register,POST /api/v1/auth/login,GET /api/v1/auth/me - Unified —
POST /api/v1/unified/export,GET /api/v1/unified/exports/{entity_id},GET /api/v1/unified/verdict/{entity_id} - Entity Memory —
GET /api/v1/memory/{entity_id}— historical assessments, drift alerts - Alerts —
GET /api/v1/alerts— Sentinel and threshold breach notifications - LLM —
POST /api/v1/llm/generate,POST /api/v1/llm/extract— direct LLM interaction for custom queries
Related — Seven Production Modules
ARIN is the central decision engine. All other SAA Alliance modules export to it for cross-domain unified verdicts: Global Risk Platform (digital twins, cascade analysis), Risk Analyzer (VaR, stress testing), Investment Dashboard (equity research), News Analytics (narrative intelligence), Crypto Analytics (digital assets), KOKON Control Company (governance, agent calibration, multi-project rollout) — see Platform Overview.