From Backtests to Agents: Building Intelligent Infrastructure for AI-Driven Finance
A systems-level approach to building AI-native infrastructure for the future of finance.
TL;DR: I started in machine learning for trading and ended up building a full-stack platform for agentic, governed, and reproducible AI workflows — from quant research to live simulation. This is how that happened.
📘 From Theory to FinRL
Before I ever touched a model, I studied the theory. I read two books on machine learning for finance, trying to understand how algorithms could learn from market data. But it wasn’t until I found a GitHub project — FinRL — that everything clicked.
That repo didn’t just teach me how the code worked. It showed me how research workflows could turn into systems. From that moment on, I wasn’t just experimenting with ML — I was building infrastructure.
🧪 Learning Through Building
My first steps were focused on learning: extending training workflows, integrating with AWS, exploring reinforcement learning through direct experimentation. I started crafting training batches, normalizing pipelines, and eventually built a private system to handle DRL-based strategy testing and iteration at scale.
As manual tweaking became a bottleneck, I added config-based experimentation to structure research contributions — a move toward reproducibility and version control.
Data friction became the next challenge, so I designed a feature engineering and serving system focused on financial workflows — now released as FinFeast.
🧠 Designing Real-World Simulation
Learning is one thing. Understanding how things work in production is another.
To gain practical insight, I joined a professional trading room led by John Carter. That gave me real-world visibility into how PMs, traders, and analysts make decisions. It completely shifted my perspective — from model accuracy to process architecture.
I began simulating not just strategies, but the people and workflows behind them. This turned into a modular system that mimics the lifecycle inside a hedge fund:
Portfolio Manager → Quant Developer → Risk Manager → Trader → Ops.
Each role is agent-driven and can be orchestrated end-to-end, including paper trading integrations and CI-tested pipelines.
🔍 From Agents to Governance
Around this time, I started exploring LLM agents and orchestration frameworks — not just to generate insights, but to build repeatable, explainable workflows.
That led to AI Architect — my open-source framework for governed LLM operations. It tracks cost, logs decisions, supports evaluation, and integrates with tools like Prometheus and Grafana. It treats agents like production systems — observable, auditable, and built for reliability.
I also created Mandala — a Spring Boot + Apache Pekko Typed (Akka) + Vue app — to deepen my understanding of enterprise backend systems and explore how these agentic workflows could plug into real infrastructure. I used it as a testing ground for integration with AI Architect.
🧾 Infrastructure That Tells the Truth
While building these intelligent systems, I ran into a very real-world problem: compliance. In Canada, SR&ED credits (R&D tax incentives) require structured evidence. So I built a local-first data lake to collect audit-ready data from Git, Trello, and Confluence — saved as Parquet, queryable with DuckDB, and privacy-aware by default.
I didn’t build it for the hype. I built it because:
Good systems tell the truth.
They log what happened, when, and why. They’re testable, observable, and reliable — whether you’re debugging a pipeline or defending an R&D claim.
🔮 Where I’m Headed
Right now, I’m pulling all these pieces together into a single direction:
A modular, agent-driven platform for quant R&D — from backtesting to live execution — with governance and auditability built in.
What I’m building:
Research → Train → Simulate → Execute → Audit
Agentic workflows for each team role (PM, Quant, Risk, etc.)
Full logging and cost tracking
Local-first with optional cloud integrations
Some components — like FinFeast and AI Architect — are already public. Others are private as I continue building toward a commercial-grade product. I’m exploring both open-core and sustainable monetization options to protect IP while keeping parts accessible.
🚀 What’s Next
Right now, I’m focused on:
• Developing the DecisionMesh architecture — refining the coordination model and exploring early validation pathways.
• Expanding BuildTheEdge — sharing frameworks, diagrams, and research notes on intelligent automation, agentic AI, and decision systems.
• Continuing the bridge between engineering and reasoning systems — exploring where practical systems design meets agentic workflows.
If you’re building in AI infrastructure, trading systems, or intelligent automation — or hiring engineers who think in systems — let’s connect.
📂 GitHub: personal & bridge-intelligence-lab
🔗 LinkedIn: linkedin.com/in/buildtheedge
📬 Email: rodrigo@bridgeintelligence.ca
Thanks for reading — more coming soon.
Thanks for reading Rodrigo Oliveira! Subscribe for free to receive new posts and support my work.