Overview
What Is This Platform?
A production-grade, cloud-native multi-agent AI platform for capital markets trade and position reconciliation — built to replace $500K–$2M/year legacy vendor licenses with an open, extensible, AI-native alternative.
The platform combines deterministic reconciliation rules, template-based decision tables, and LLM reasoning into a three-layer resolution engine. Five Model Context Protocol (MCP) servers expose data, file, cache, and computation capabilities to a set of seven specialised AI agents. The result is a fully automated reconciliation workflow that handles the full break lifecycle from detection through root-cause analysis to resolution narrative generation.
434
Passing Tests
5
MCP Servers
7
AI Agents
$2M
Vendor License
Replacement Value
Replacement Value
Primary Value Driver: The platform's lead strategic positioning is vendor license replacement — not headcount reduction. Legacy reconciliation platforms (Intellimatch, TLM, Duco equivalents) typically cost $500K–$2M/year in license fees. This platform replaces that cost with an open-source, AI-native architecture that also adds LLM-powered break resolution and regulatory narrative generation that legacy platforms cannot provide.
Architecture
System Architecture
The platform follows a clean layered architecture: data sources at the bottom, MCP infrastructure in the middle, agent orchestration above, and the resolution engine at the top. The architecture is 12-Factor compliant and designed for cloud-native deployment.
Layer 1 — Orchestration & Workflow
Bootstrap Agent (ML pre-training), Reconciliation Orchestrator, Break Resolution Manager. Coordinates the full break lifecycle from ingestion to closure.
Bootstrap Agent
Orchestrator
Break Manager
↓
Layer 2 — Resolution Engine (Three Layers)
Layer 1: Deterministic rules (exact match, tolerance matching, known-break patterns). Layer 2: Template and decision table matching (pre-trained ML classification). Layer 3: LLM reasoning for novel breaks requiring contextual analysis.
Deterministic Rules
Decision Tables
ML Classifier
LLM Reasoning
↓
Layer 3 — MCP Server Infrastructure
Five MCP servers expose tools and resources to the agent layer via the Model Context Protocol. Each server runs as an independent microservice with its own port and data domain.
SQLDB :8083
FILE :8084
CACHE/Redis :8085
Analytics MCP
Reporting MCP
↓
Layer 4 — Data & Persistence
Five supported database types (PostgreSQL, MySQL, SQLite, SQL Server, Oracle). Redis caching layer for break state management and session persistence. File-based data ingestion supporting industry-standard formats.
PostgreSQL
MySQL
SQLite
SQL Server
Oracle
Redis Cache
↓
Layer 5 — License & Deployment
14-day trial enforcement system with HMAC-signed license files. License generator tool. Cloud-native deployment configurations. 12-Factor compliant application design. GPU/CUDA support planned for future ML workloads.
HMAC License Enforcement
14-Day Trial
Docker
Cloud-Native
Agent Roster
Seven Specialised AI Agents
| Agent | Primary Responsibility | MCP Tools Used | Output |
|---|---|---|---|
| Bootstrap Agent | ML model pre-training on historical break patterns before live deployment | SQLDB, FILE, Analytics | Trained classification models |
| Ingestion Agent | Data validation, format normalisation, source reconciliation setup | FILE, SQLDB | Normalised position/trade data |
| Break Detection Agent | Identify and classify breaks across trade, position, and cash records | SQLDB, Analytics | Break records with severity score |
| Resolution Agent | Three-layer break resolution (deterministic → ML → LLM) | All MCP servers | Resolution recommendation + confidence |
| Escalation Agent | Break escalation routing, SLA monitoring, priority queue management | CACHE, Reporting | Escalation notifications + routing |
| Narrative Agent | LLM-powered break explanation and regulatory commentary generation | Analytics, Reporting | Human-readable break narratives |
| Audit Agent | Full audit trail, reconciliation sign-off, regulatory reporting | SQLDB, FILE, Reporting | Signed audit records + reports |
Value Proposition
Why This Architecture Matters
Primary Value Driver
Vendor License Replacement
Legacy reconciliation platforms (Intellimatch, TLM, SmartStream equivalents) carry $500K–$2M/year in license fees. This platform replaces that spend with an open-source architecture while adding AI capabilities those legacy platforms cannot provide.
AI Capability Gap
LLM Break Reasoning
Legacy platforms apply deterministic rules only. This platform adds a third resolution layer — LLM reasoning for novel, contextual, or cross-system breaks that rule-based engines systematically fail on. Break resolution rate improves as the ML layer trains on institutional patterns.
Regulatory Value
Audit Trail & Narrative
Every break resolution is logged with full chain-of-thought reasoning. The Narrative Agent produces human-readable explanations that satisfy regulatory audit requirements. Compliance teams get documentation that supports regulatory examination without manual writeup.
Architecture Quality
Production-Grade Build
434 passing tests across 193 files. 12-Factor compliant. Five supported database types. HMAC-signed license enforcement. Cloud-native deployment configs. GPU/CUDA roadmap for ML acceleration. This is not a proof-of-concept — it is a deployable platform.
Portfolio Significance: This project demonstrates the full CCA exam domain coverage in practice: Agentic Architecture and Orchestration (7 agents, multi-step workflows), Claude Code Configuration (30 commits, 193 files), Prompt Engineering for reconciliation context, Tool Design and MCP Integration (5 MCP servers), and Context Management for long break resolution chains.
Technical Details
Implementation Notes
Three-Layer Resolution Engine
The resolution engine is the core intellectual contribution of the platform. Break resolution proceeds through three layers in sequence, with escalation between layers based on confidence scores:
- Layer 1 — Deterministic Rules: Exact match, tolerance-based matching (configurable thresholds per asset class and break type), known-pattern matching against a curated rule library. Resolves approximately 60–70% of breaks in a typical institutional environment. Zero LLM token consumption.
- Layer 2 — Template and Decision Table: ML classification model pre-trained by the Bootstrap Agent on historical break data. Applies learned patterns to classify novel breaks against known categories. Handles approximately 20–25% of remaining breaks. Confidence threshold configurable per deployment.
- Layer 3 — LLM Reasoning: Claude API-powered contextual analysis for breaks that Layer 1 and Layer 2 cannot resolve with sufficient confidence. Provides detailed break explanation, likely root cause, and recommended resolution action. All LLM calls are tool-augmented via MCP servers for data retrieval — the LLM never operates on cached context alone.
# Simplified resolution flow
result = deterministic_rules.resolve(break_record)
# confidence > 0.95 → auto-resolve
if result.confidence < 0.95:
result = ml_classifier.classify(break_record)
# confidence > 0.80 → auto-resolve with audit
if result.confidence < 0.80:
result = llm_agent.reason(
break_record,
tools=[sqldb_mcp, file_mcp, analytics_mcp]
)
# human review required if LLM confidence < 0.70
MCP Server Design
Each MCP server is designed around a specific data domain and exposes tools following the Model Context Protocol specification:
| MCP Server | Port | Tools Exposed | Primary Use |
|---|---|---|---|
| SQLDB Server | :8083 | query, execute, schema, explain | Trade & position data queries |
| FILE Server | :8084 | read, write, list, diff, archive | Reconciliation file I/O |
| CACHE/Redis | :8085 | get, set, delete, scan, expire | Break state & session mgmt |
| Analytics MCP | :8086 | aggregate, classify, cluster, score | Break pattern analysis |
| Reporting MCP | :8087 | generate, export, sign, distribute | Audit trail & regulatory reports |
12-Factor Compliance: The platform follows Heroku's 12-Factor App methodology throughout: configuration via environment variables, stateless processes, port-bound services, log streaming to stdout, and disposable container design. This ensures seamless deployment to any cloud provider (AWS, Azure, GCP) or on-premise Kubernetes environment without code changes.
Roadmap
Planned Extensions
- GPU/CUDA acceleration — ML model inference acceleration for high-volume batch reconciliation windows using CUDA-enabled containers.
- Digital asset support — blockchain transaction reconciliation with on-chain data verification via DeFi protocol adapters.
- Private license repository — separate private repo for license generation artifacts to enable commercial deployment.
- Local demo environment — packaged Docker Compose demo with pre-seeded break scenarios for rapid evaluation by prospective clients.
- Fractal risk integration — integration with the Risk Intelligence fractal engine for risk-adjusted break prioritisation and capital impact estimation.
CAPITAL MARKETS RECONCILIATION PLATFORM
Multi-Agent · MCP-Native · AI-Powered · Production-Grade
30 commits · 193 files · 434 passing tests · 5 MCP servers · 7 agents · Cloud-native