Build compliant investment portfolios with AI risk agent

Copy link

AI multi-agent system that automates portfolio construction, compliance checking, and risk analysis for investment advisors

Table of contents

Copy link

1. Detailed description
   • The Challenge
   • Our Solution
   • Our Solution Stack
   • Architecture diagrams
2. Requirements
   • Minimum hardware requirements
   • Minimum software requirements
   • Required user permissions
3. Deploy
   • Quick Start - OpenShift Deployment
   • Quick Start - Local Development
   • Usage
   • Delete
4. References
5. Tags

Detailed description

Copy link

The Challenge

Copy link

Portfolio managers at asset management firms must build investment portfolios that satisfy each client's unique guidelines, meet risk tolerances, avoid prohibited securities, and hit return targets. Now multiply that across tens or hundreds of clients. The manual process — reading guideline documents, cross-referencing prohibited tickers, running risk calculations, and drafting client communications — is slow, error-prone, and does not scale.

Our Solution

Copy link

An AI multi-agent system that automates the end-to-end portfolio construction workflow in two phases:

Phase 1 — Deterministic Pipeline:

• Parses client investment guideline PDFs and extracts prohibited ticker symbols using an MLP classifier
• Builds a compliant equal-weight equity portfolio from the S&P 100, excluding prohibited securities
• Calculates 1-day parametric Value at Risk (VaR) at 99% confidence
• Automatically retries portfolio construction if VaR exceeds the client's risk threshold (up to 10 attempts)
• Drafts a client-ready summary email via the LLM

Phase 2 — Agentic Chat:

• Interactive chat where the LLM can call portfolio and VaR tools to modify holdings and recalculate risk
• Prohibited tickers from Phase 1 persist — the advisor cannot violate compliance constraints
• VaR is automatically recalculated after every portfolio change

Our Solution Stack

Copy link

AI/ML

Copy link

• LLM — powers email drafting, agentic chat reasoning, and tool selection (served via OpenShift AI model serving or any OpenAI-compatible endpoint). Tested with Llama 3.3 70B quantized.
• scikit-learn MLP — classifies guideline sentences as prohibition-related to extract banned tickers.

Backend Services

Copy link

• Flask — REST API orchestrator + 3 specialized tool agents.
• OpenAI Python SDK — LLM function calling for agentic chat.
• yfinance — real-time market data for portfolio pricing and VaR calculation.

Frontend

Copy link

• React + Vite + TypeScript — two-tab UI for pipeline execution and portfolio chat.

Infrastructure

Copy link

• Podman/Docker Compose — local development.
• Red Hat OpenShift + Helm — production deployment.
• Optional: Knative — serverless agent auto-scaling.

Red Hat® OpenShift® AI

Copy link

• Model Serving (KServe) — serves the scikit-learn MLP classifier as an InferenceService (MLServer sklearn runtime, S3/MinIO storage).
• NeMo Guardrails (TrustyAI) — input/output filtering that blocks sensitive data and off-topic requests.

Architecture diagrams

Copy link

Phase 1: Deterministic Pipeline

Copy link

The UI calls granular orchestrator endpoints in sequence, reimplementing the retry loop client-side for real-time progress feedback.

Figure 1: Deterministic pipeline — parse guidelines, build portfolio, calculate VaR with retry loop, draft email.

flowchart TD
  Start["User submits pipeline request"] --> G["1. Parse Guidelines PDF"]
  G -->|"prohibited tickers"| P["2. Build Portfolio"]
  P -->|"holdings"| V["3. Calculate VaR"]
  V --> Check{"VaR > max?"}
  Check -->|"Yes, up to 10x"| P
  Check -->|"No"| E["4. Draft Client Email via LLM"]
  E --> Done["Return results"]

Phase 2: Agentic Chat

Copy link

The LLM selects from portfolio_equities, portfolio_replace_symbol, and value_at_risk tools. Guidelines are frozen — prohibited tickers cannot be overridden. VaR is recalculated automatically after every portfolio mutation.

Figure 2: Agentic chat — LLM selects tools to modify holdings and recalculate VaR while respecting frozen guidelines.

sequenceDiagram
  participant User
  participant UI
  participant Orchestrator
  participant LLM
  participant Tools as Portfolio/VaR Agents

  User->>UI: Chat message
  UI->>Orchestrator: POST /chat with message + context
  Orchestrator->>LLM: System prompt + tools
  LLM-->>Orchestrator: Tool call
  Orchestrator->>Tools: POST /tools/tool_name
  Tools-->>Orchestrator: Result
  Orchestrator->>LLM: Tool result
  LLM-->>Orchestrator: Final response
  Orchestrator-->>UI: Response + updated context
  UI-->>User: Display answer + refreshed portfolio outputs

See it in action

Copy link

Deploy locally with make deploy-local and open the UI at http://localhost:8080, or deploy to OpenShift with make deploy-cluster and open the route (see Usage). Explore the pipeline and chat tabs.

Requirements

Copy link

NOTE: This quickstart assumes a large language model is already deployed in your environment. Guidance for deploying models is available in the References section.

Minimum hardware requirements

Copy link

Application:

• CPU: 2 vCPU (request) / 4 vCPU (limit)
• Memory: 4 GiB (request) / 8 GiB (limit)
• Storage: 10 GiB

• Optional: GPU — required if you plan to deploy your own model on OpenShift AI, for deployment refer to References.

Minimum software requirements

Copy link

OpenShift Cluster Deployment

Copy link

• Red Hat OpenShift AI 2.25.0 or later (for model serving)
• Helm 3.0.0 or later
• oc CLI (for OpenShift)
• LLM Model Server — Red Hat OpenShift AI model serving (vLLM) recommended, or any OpenAI-compatible endpoint.

Local Development Requirements

Copy link

• Podman and Podman Compose (or Docker and Docker Compose)
• Make (for running deployment commands)

Required user permissions

Copy link

Local Development

Copy link

• Permission to run containers via Podman or Docker
• Access to local ports: 5000, 7001, 7002, 7003, 8080

OpenShift Cluster Deployment

Copy link

• Authenticated to the OpenShift cluster (oc login)
• Permission to create or access a namespace/project

Deploy

Copy link

Quick Start - OpenShift Deployment

Copy link

1. Clone the repository:

git clone https://github.com/rh-ai-quickstart/portfolio-manager-agent.git
cd portfolio-manager-agent

2. Deploy to OpenShift:

make deploy-cluster

3. Get the UI URL:

oc get route ui -n portfolio-manager-agent -o jsonpath='https://{.spec.host}'

This reads your .env file (if present) and passes any non-empty OPENAI_* values to Helm, so the UI starts with LLM connection settings pre-filled.

To deploy with serverless (Knative) agents:

helm upgrade --install portfolio-manager-agent deploy/helm \
  -n portfolio-manager-agent --create-namespace \
  --set serverless.enabled=true

Quick Start - Local Development

Copy link

1. Clone and Setup Repository

Copy link

git clone https://github.com/rh-ai-quickstart/portfolio-manager-agent.git
cd portfolio-manager-agent

2. Configure Environment Variables

Copy link

cp .env.example .env

Edit .env with your LLM configuration:

OPENAI_API_ENDPOINT=https://your-llm-host/v1   # OpenAI-compatible endpoint (include /v1)
OPENAI_API_TOKEN=sk-your-api-key                # API key for the endpoint
OPENAI_MODEL=llama-3-3-70b-instruct-w8a8        # Model name

Tip: For enterprise deployments, we recommend using Red Hat OpenShift AI model serving to host your LLM with built-in GPU support, autoscaling, and enterprise security.

3. Start All Services

Copy link

make deploy-local
# equivalent to: podman compose -f deploy/local/compose.yml up -d --build

Usage

Copy link

Open the UI at http://localhost:8080 (local) or the OpenShift route:

oc get route ui -n portfolio-manager-agent -o jsonpath='https://{.spec.host}'

If you created a .env file, the LLM endpoint, API key, and model are loaded automatically. Otherwise, expand Connection settings to enter them manually.

Tab 1 — Portfolio Setup:

1. Set the Investment guidelines URL, Portfolio value, Number of symbols, and Max VaR.
2. Click Run pipeline. A live progress log shows each step: guidelines parsing, portfolio construction (with VaR retry loop), and email drafting.
3. On success, results appear in Portfolio outputs and the chat tab unlocks.

Tab 2 — Discuss Portfolio:

• Chat about the portfolio — the LLM can call tools to swap holdings, rebuild the portfolio, or recalculate VaR.
• Portfolio outputs refresh automatically when the context changes.

Delete

Copy link

Stop Local Deployment

Copy link

podman compose -f deploy/local/compose.yml down

Delete from OpenShift

Copy link

helm uninstall portfolio-manager-agent -n portfolio-manager-agent

References

Copy link

• How to deploy language models with Red Hat OpenShift AI