I am not a software engineer. I have never written production code in my life.
Over recent months, I built a functioning AI-native macro research platform designed around institutional investment workflows — a weekly regime classification engine, automated data pipelines, a structured investment universe database, and a discovery engine that finds cross-theme convergence signals. It runs automatically every Monday morning before the markets open. It emails me a report, updates the website, and generates a structured research agenda for the week.
I did not hand this to a technology team. What I did was something different — and what I learned in the process is, I think, worth sharing.
The Architecture Most People Miss
Most discussions of "building with AI" focus on the language model and ignore the layer that makes sustained, compounding work possible. The system I built has two distinct components.
Claude (Anthropic) is the reasoning layer — writing code, thinking through architecture, drafting analysis, and pushing back on assumptions when they deserve to be questioned. It works at a pace no human developer could match, holding the entire codebase in working memory simultaneously.
Hermes is the persistence and automation layer — an open-source AI agent framework that provides everything a language model lacks: persistent memory across sessions, scheduled automation on a cron schedule, file access, a Telegram interface, and deployment tooling that allows the system to act on the world rather than simply describe it. Earlier generations of agent frameworks — such as OpenClaw, the harness developed by Peter Steinberger — pioneered this approach. Hermes is the framework I use today, and it is what transforms Claude's reasoning from a conversation into a compounding research platform.
Without an agent layer, AI-assisted research is brilliant in bursts and forgetful between them. With one, it accumulates. That is the entire difference between a clever tool and a compounding research platform.
Three Collaborators, One Process
What emerged — not by design at the outset, but through the reality of extended use — was a collaboration between three distinct AI capabilities, each playing a role the others could not.
Claude and Hermes handled the building and operating. ChatGPT (OpenAI) became the strategic layer — and this is the role most articles about AI-assisted work omit.
ChatGPT acted as an independent strategic reviewer, challenging assumptions, testing whether architectural decisions actually improved investment outcomes, and applying consistent pressure against one question: Will this actually improve investment decisions?
When I was deep in the process of building — adding scoring dimensions, designing classification layers, extending the database schema — it was easy to confuse intellectual elegance with genuine usefulness. ChatGPT was the collaborator that consistently asked whether any of it changed a decision. When the honest answer was no, it said so plainly.
When I proposed adding a geopolitical stress indicator as a ninth category in my macro regime engine, the pushback was immediate: that belongs in the theme definition layer, not the regime engine. The regime engine measures financial conditions. Geopolitical signals belong in the analytical layer above it. Conflating them would corrupt the primary instrument. The design decision was reversed. The platform is better for it.
The three capabilities are not redundant. They are complementary. Claude excelled at building. Hermes excelled at operating. ChatGPT excelled at challenging. And behind all three: the human investor providing the judgement that none of them could supply.
What AI's Real Advantage Is Not
You will read many claims about AI replacing analysts and researchers. Some of those claims will prove true over time. But they miss the more immediate and more interesting point.
AI's real advantage — at this moment, in the way I have used it — is not replacement. It is cycle compression.
The gap between "I wonder if this scoring dimension actually drives investment decisions" and "I have analysed the data and confirmed it" used to be measured in days. In my workflow, that analysis happened in a single conversation. The time from idea to evidence was under a minute.
Compress that cycle hundreds of times, and the compound effect is not incremental. It is transformational. The judgement that determines which hypotheses are worth testing — that remains mine. AI accelerated every stage of the process. It did not remove my responsibility for deciding what was worth believing.
The Architecture of Discovery
Early in the build, I made a decision I now regard as the most important one in the entire project.
I was building a platform designed to find investment opportunities I had not explicitly anticipated — vehicles sitting at the intersection of multiple independent analytical frameworks simultaneously. The temptation, when building something like this, is to design the intersections you want to find.
I resisted that. The discipline I imposed: each investment theme is seeded by an independent analytical session, with no knowledge of what other sessions have done. The convergence engine — and only the convergence engine — is permitted to look across themes.
A finding counts as a discovery only if the two theme memberships that create the intersection were established independently.
That sounds like a small design decision. In practice, it determines whether the platform discovers ideas or merely confirms existing beliefs.
The test case: Palantir Technologies. I added Palantir to the AI Infrastructure universe in one analytical session, as a data analytics platform. Several sessions later, building the Defence & Aerospace universe, I independently tagged Palantir as a primary expression of the defence intelligence theme. Neither session referenced the other. When the convergence engine ran, two independent lines of analytical reasoning had arrived at the same vehicle.
The platform did not tell me to buy Palantir. It told me that two independent investment theses deserved to be considered together. The investment decision remained mine. Large investment organisations occasionally discover this kind of convergence when analysts working independently arrive at the same company from different directions. I designed a platform that makes those convergences systematic rather than incidental.
The Compound Architecture
The final insight is about time. Most analytical processes are flat. Each new analysis starts from scratch. The reasoning that led to a view is not systematically captured. The condition that would change the view is not documented.
The platform I have built is designed to be different. Every finding is logged with its provenance. Every decision is recorded with three fields: the question that was asked, the decision that was made, and the condition that would cause the decision to be revisited.
Every rejected investment idea is documented with a specific revisit trigger — not "I didn't like it" but "I would reconsider if the valuation compresses to 40 times next twelve months' earnings." That is a precise instruction to a future version of the same analytical engine.
In thirty years of investment management, I never had a research tool that remembered what I had decided and why, and could tell me when the conditions I had specified had changed. Now I do.
What This Means
Some of the analytical advantages that once depended on institutional scale are now becoming available to much smaller organisations willing to build the right research infrastructure.
The building layer provides the implementation capacity. The operating layer provides the persistence and automation. The strategic layer provides the challenge that keeps complexity honest. The human provides the judgement that none of them can supply.
The disciplines that made it work — independent seeding, structured decision records, annotated rejections, the integrity to ask whether complexity serves the purpose or merely satisfies the builder — are available to any practitioner willing to apply them.