AlphaForge AI — An Investment Analyst That Can Say No, and Show Why
Inspiration
Investors have access to more financial information than ever, yet understanding whether an investment is appropriate for their personal circumstances remains difficult.
Traditional platforms display returns, volatility, and risk scores, but often fail to explain what those numbers actually mean for an individual investor.
Meanwhile, generative AI introduces another problem: a chatbot can confidently recommend an investment, invent supporting statistics, or overlook important financial constraints.
We wanted to build something different.
What if AI could investigate investment decisions like a financial analyst, while deterministic systems ensured that every recommendation respected the investor's constraints?
That question inspired AlphaForge AI.
What It Does
AlphaForge AI is a multi-agent investment analysis and simulation platform designed to make portfolio decisions more transparent, personalized, and explainable.
Users can describe their financial situation, complete a guided questionnaire, or submit an existing portfolio for analysis.
The platform then:
- Builds a financial profile: Evaluates investment horizon, financial capacity, risk tolerance, financial knowledge, and liquidity requirements.
- Calculates personalized constraints: Determines loss limits, diversification requirements, defensive allocations, and individual asset exposure limits.
- Generates portfolio strategies: Constructs up to three simulated portfolios with different risk levels, using historical market data and quantitative risk models.
- Validates every strategy: A deterministic Guardian checks suitability, concentration, diversification, losses, and consistency before presenting results.
- Explains the reasoning: AI agents investigate potential weaknesses and explain the findings in accessible language.
- Supports interactive scenarios: Users can challenge recommendations and explore alternative allocations through conversational AI.
A Concrete Example
Consider Robert, a fictional 59-year-old investor with a $400,000 portfolio and a maximum acceptable loss of 15%.
Robert wants to allocate 25% of his portfolio to Tesla.
Using historical market data, AlphaForge identifies a worst observed Tesla drawdown of 70.9%. At Robert's proposed allocation, that decline alone would represent approximately 17.7% of his portfolio, exceeding his entire loss tolerance.
Rather than blindly approving his request, AlphaForge explains the conflict and calculates an allocation of approximately 3.1% that fits its modelled constraints.
The goal is not to make decisions for investors, but to help them understand the consequences before deciding.
How We Built It
AlphaForge AI uses a hybrid architecture combining agentic artificial intelligence with deterministic quantitative finance.
Multi-Agent Architecture
We implemented four specialized agents using the AWS Strands Agents SDK, powered by Anthropic Claude Opus 5.5 through Amazon Bedrock.
- Extractor: Converts natural-language financial descriptions into structured profiles while checking that numerical inputs originate from the user.
- Critic: Investigates proposed portfolios through quantitative tools, identifies weaknesses, and suggests constrained revisions.
- Explainer: Generates accessible explanations grounded in verified financial calculations.
- Analyst: Handles follow-up questions and portfolio what-if scenarios by calling quantitative tools and rerunning the analysis pipeline when necessary.
A deterministic Python pipeline orchestrates the workflow and ensures that mandatory validation steps cannot be bypassed.
Quantitative Risk Engine
Our financial engine analyzes historical market data for 31 assets, converted into Canadian dollars, primarily covering 2021–2026.
It incorporates portfolio drawdowns, volatility, correlations, diversification, ETF look-through, historical stress scenarios, and bootstrap-based risk simulations.
Portfolio allocations are constrained by the investor's circumstances rather than being generated freely by a language model.
The Guardian: AI With Enforceable Boundaries
At the center of AlphaForge is a deterministic Guardian implementing 17 validation rules, including checks inspired by Canadian know-your-client and suitability principles.
The Guardian verifies proposed strategies, checks revisions, and validates AI explanations.
One particularly important mechanism prevents the Explainer from inventing financial figures: the model writes numerical placeholders that are replaced with values computed by the financial engine.
AI investigates and explains. Code calculates and enforces.
AWS Integration
Amazon Bedrock powers the specialized AI agents, while Strands provides their tool-calling capabilities.
Amazon S3 stores analysis traces and supports access to the application's datasets, enabling the team to inspect calculations, agent actions, and validation results.
Optional voice interaction integrates Amazon Transcribe and Amazon Polly.
The user-facing application is built with Python and Streamlit, with French and English interfaces.
Challenges We Faced
Our biggest challenge was making generative AI useful in a financial setting without allowing it to invent numbers or override risk constraints.
We addressed this through strict tool grounding, deterministic calculations, mandatory Guardian checks, and fallback mechanisms.
Another challenge involved portfolio optimization. We experimented with mean-variance optimization and dynamic portfolio protection, but found that more sophisticated approaches did not necessarily produce more reliable or understandable results. We prioritized transparent, constraint-driven strategies instead.
We also had to distinguish historical observations from future risk estimates, documenting the assumptions and limitations of our simulation model.
Accomplishments We're Proud Of
- Implementing a four-agent investment analysis workflow with mandatory deterministic validation.
- Creating personalized financial constraints instead of relying solely on generic investor categories.
- Supporting interactive portfolio scenarios with full recalculation and validation.
- Building an auditable system that records the evidence behind its outputs.
- Developing an offline evaluation suite with 16 scenario profiles and numerical-grounding tests.
- Providing bilingual interfaces and a guided experience designed for non-expert investors.
What We Learned
This project reinforced an important principle: the strongest financial AI systems are not necessarily those that give AI the most autonomy, but those that combine intelligent reasoning with verifiable constraints.
We learned how to design tool-using agents, structure quantitative financial workflows, evaluate AI-generated claims, and integrate multiple AWS services into an auditable application.
What's Next
Future work includes improving risk-model calibration, expanding historical validation, strengthening portfolio optimization, and incorporating more comprehensive financial planning constraints.
We also want to explore how this architecture could support human financial advisors by making investment analysis more transparent and easier to audit.
Our vision is an investment assistant that does more than provide answers: one that can explain its reasoning, recognize its limitations, and show the evidence behind every decision.
AlphaForge AI is an educational simulation using fictional profiles. It does not provide financial advice or execute real transactions.
Built With
- altair
- amazon-bedrock
- amazon-polly
- amazon-transcribe
- amazon-web-services
- anthropic-claude
- numpy
- pandas
- python
- strands-agents-sdk
- streamlit