Airia Sentinel

Inspiration

Treasury teams at mid-to-large corporations face a daily challenge: monitoring dozens of financial instruments across crypto, forex, and commodities while managing corporate currency exposure. Manual processes are slow, error-prone, and cannot scale. We asked: what if 10 specialized AI agents could collaborate in real-time to deliver institutional-grade risk analysis in under 10 seconds?

What it does

Airia Sentinel is a multi-agent orchestration platform for treasury risk management. It runs a 5-phase pipeline of 10 AI agents that work together to:

Scan markets — Real-time data from 10 financial instruments (BTC, ETH, SOL, EUR/USD, XAU/USD, etc.)
Detect anomalies — Z-score analysis, volume spikes, divergence detection, contagion alerts
Score sentiment — Multi-factor sentiment analysis with AI classification
Aggregate risk — Weighted fusion of all signals with cascade detection
Build consensus — 3-way AI consensus (LM Studio M1 + M2 + Ollama + Airia cloud)
Size positions — Kelly Criterion + Risk Parity optimal allocation
Backtest strategies — Monte Carlo simulation (50 trajectories x 30 days, Sharpe/MaxDD/Calmar)
Check compliance — Automated regulatory compliance via Airia
Generate reports — 6-section executive report with Airia narrative enrichment
Dispatch alerts — 4-severity alert system (INFO/WARNING/CRITICAL/EMERGENCY)

A real-time web dashboard displays the full pipeline: agent status, market signals, risk gauges, consensus strategies, cluster health, and audit trail. Human-in-the-Loop (HITL) approval gates ensure no strategy executes without human validation.

How we built it

Architecture

Phase 1 (parallel)    Phase 2        Phase 3 (parallel)   Phase 4        Phase 5 (parallel)
┌────────────┐       ┌──────────┐   ┌──────────────┐     ┌──────────┐   ┌────────────┐
│ Market     │──┐    │ Risk     │──►│ Consensus    │     │ Strategy │──►│ Compliance │
│ Intelligence│  │──►│Aggregator│   │ Strategy     │     │Backtester│   │ Report     │
│ Corporate  │──┘    │          │──►│ Position     │──►  │          │   │ Alert      │
│ Anomaly    │       └──────────┘   │ Sizer        │     └──────────┘   └────────────┘
│ Sentiment  │                      └──────────────┘
└────────────┘

Tech Stack

Airia SDK — Cloud AI pipelines for market analysis, compliance, consensus, backtesting, and report synthesis (10 bilingual prompts EN/FR)
LM Studio cluster — 3 nodes (M1: qwen3-30b on 6 GPUs, M2: deepseek-coder on 3 GPUs, M3: mistral-7b)
Ollama — Local lightweight inference (qwen3:1.7b) for fast sentiment and routing
Python asyncio — Parallel agent execution with asyncio.gather()
FastAPI + WebSocket — Real-time dashboard server with live pipeline events
SQLite — Full audit trail with WAL mode for concurrent access
NumPy — Monte Carlo GBM simulation, Kelly Criterion, Risk Parity calculations
Rich — Terminal UI with colored tables, panels, and progress indicators
CCXT — Multi-exchange market data (MEXC Futures)

Key Design Decisions

Hybrid AI: Every agent tries Airia cloud first, then falls back to local LM Studio/Ollama — ensuring the platform works even offline
Bilingual prompts: All 10 Airia prompts have FR and EN versions with a single PROMPT_LANG toggle
Pipeline Engine: Supports 3 orchestration patterns (Domino sequential, Vectorial parallel, Matrix grid scoring)

Challenges we ran into

LM Studio auth: API key headers were missing on cluster health checks and query calls — caused silent failures until we traced it with audit logs
Monte Carlo reproducibility: Needed deterministic seeds per strategy (hash(run_id + strategy_name)) while keeping simulations statistically meaningful
10-agent coordination: Passing data between 5 pipeline phases required careful schema design — we standardized on StepResult dataclass with data dict, confidence, latency_ms
Dashboard without React: Built a 1800-line single-file HTML dashboard with vanilla JS, SVG gauges, and CSS animations to avoid build tooling complexity

Accomplishments that we're proud of

10 agents, 5 phases, under 10 seconds — Full pipeline from market scan to executive report
Zero-dependency dashboard — Single HTML file with live WebSocket updates, no npm/build required
Monte Carlo backtester — 50 GBM simulations with Sharpe, MaxDD, Calmar metrics and letter grades
True multi-model consensus — 4 different AI providers vote on strategy with weighted scoring
Human-in-the-Loop — Real approve/reject buttons that gate strategy execution

What we learned

Airia SDK is remarkably flexible — the execute_pipeline() and execute_temporary_assistant() APIs let us integrate cloud AI at every pipeline stage while keeping local fallbacks
Multi-agent systems need strong contracts between agents — our StepResult pattern made it trivial to add new agents
Real-time dashboards massively improve the "Design" dimension — seeing agents complete in sequence is more convincing than any log file