Hiring Twenty Analysts in a Closet
A Continuation Update on the Solo Hedge Fund Architecture
When I wrote about building a solo hedge fund last time, the architecture was still a sketch — a few launchd jobs, one persona, a vault of unstructured DD notes, and a fragile mental model of how everything connected. Most of what worked, worked only because I was the one watching it. Most of what was supposed to work didn't, because I was the only one watching it.
Since then, the architecture has evolved. Today I want to share how the architecture is shaped, what it makes possible that wasn't possible before, and where the human judgment lives — because it absolutely still lives somewhere, and recognising that is the whole game.
The team you can't fire
The mental model I keep coming back to is: I am running a small investment team. Not in the metaphorical sense — in the literal, organisational-chart sense. There is a portfolio manager (me). There are sector analysts. There is a risk officer. There is a macro overlay. There is a valuation specialist. There is a forensic short-seller. There is a position-sizing specialist. There is a decision-quality auditor. There are post-mortem reviewers and a trade-execution log keeper. There are monitoring analysts who watch existing positions and alert me when something deserves my attention.
Each of them has a distinct voice, a distinct frame, a documented set of mental models, and — this is the part that took me longest to get right — a documented set of blind spots. The risk officer doesn't size positions. The sizing specialist doesn't kill thesis. The valuation specialist doesn't tell stories about narrative. They are not interchangeable, and I do not let them collapse into a single generic voice.
When I'm thinking about a new name, I can convene a roundtable. They debate. They disagree. They reach convergence or surface impasse. I read the transcript, decide which arguments survive cross-fire, and make a call. The call is mine. The debate is theirs.
This was not possible for a single person before. Not because the personas didn't exist — Munger and Marks and Mauboussin and Druckenmiller have all been writing publicly for decades — but because operationalizing the personas in real time, against a specific company, on a specific day, with the specific evidence on the table, required a team. Or it required an extraordinary memory and an unrealistic amount of personal time. The agentic team makes the operationalisation cheap.
What the architecture does, in big strokes
I won't walk you through file structures or pipeline DAGs. The big strokes are these:
A staged forward model pipeline. Every stock I underwrite goes through a sequence: raw inputs are assembled, conviction deltas are derived from differentiated views, valuation anchors are computed, scenarios are built, the math is checked, the assumptions are challenged. Each stage has gating criteria. Each stage produces an artifact I can audit. The pipeline is not the analysis. The pipeline is the plumbing that makes the analysis honest.
An IC roundtable layer above it. When the pipeline finishes producing a model, the multi-persona roundtable convenes to argue about the assumptions baked into it. The argument is structured by mental-model lens — fragility, base rates, narrative, sizing, valuation, governance, moat. Each persona has 400 words to make a case. The judge synthesises. The output is a structured verdict with directives I can act on, not a vibe.
A re-underwriting cycle. When new evidence lands — a quarterly print, a regulatory ruling, a competitor announcement, a change in cycle position — the position is re-underwritten through a 3-step process. The steps blend into a weighted view that I review at the layer of *individual factors* — not at the level of "is the stock a buy". The output is whether each factor still holds, whether thesis is still intact, whether kill switches have been triggered, whether the sizing is still right. Position changes follow the factor changes, not the price chart.
Continuous monitoring with alert hygiene. A separate set of agents watches my existing book — earnings transcripts, alt-data signals, sell-side downgrades, peer-group sentiment, regulator filings. They emit signals when the evidence accumulates past a threshold. The threshold is set so the alerts I get are signal-dense; the alerts I don't get are not noise I missed but noise I correctly filtered.
Post-mortem feedback loops. When a position resolves badly, I run a post-mortem. The post-mortem produces a generalizable pattern, not a case-specific lesson. The pattern gets folded back into the relevant analyst's framework as a worked example. The next time the analysts and IC encounters a structurally similar setup, the pattern is in scope. The team is learning.
That last one is the piece I treasure the most. Most investment processes treat post-mortems as documents that get archived. The post-mortems here become upgrades to the team's skill set. Compounding learning across cases, not just within them.
What was previously impossible
The solo PM constraint, before agentic AI, was brutal. You had:
Sourcing capacity bottleneck: roughly the volume of company filings, sell-side decks, transcripts, alt-data, and news flow that a single human can read in a day. Generously, 5-10% of what an institutional team covers.
Lack of perspective diversity: roughly the number of mental models you can hold simultaneously while reading the same source material. Generously, 2-3 before fatigue kicks in.
Monitoring throughput bottleneck: roughly the number of existing positions you can watch closely without missing inflection. Generously, 5-10 names if you want to know them deeply.
Memory bottleneck: roughly the number of past lessons you can retrieve at the moment they're relevant. Generously, the ones you've recently written about.
Time constraints: face it, we all need a full time day job to pay off the utility bills. Investing in stocks in oftentimes a hobby.
Each of these constraints is now an order of magnitude relaxed. Sourcing volume scales with API calls and not human attention. Perspective diversity scales with the number of distinct personas I'm willing to maintain. Monitoring throughput scales with the number of cron jobs I'm willing to schedule. Memory becomes a first-class artifact — every lesson the team learns gets retrieved at the moment of relevance, not by accident of recall.
This isn't AI-replaces-PM. This is solo-PM-now-has-the-leverage-of-a-team.
Where the human judgment lives
Here is the core observation, and it is the one most agentic-AI commentary gets wrong.
Human judgment stands at the centre of everything. Every architecture decision — what gating criteria to require, which personas to seat, how to weight the arms, what to monitor and what to ignore, when to override the team and act on conviction — is mine. The investment judgment at the moment of deployment is mine. The decision of which directive to act on and which to defer is mine. The decision of which post-mortem pattern is generalizable and which is case-specific is mine.
What the agent team does is make my judgment scalable. It sources information at a volume I could not source. It surfaces patterns I could not surface alone. It runs the same disciplined process across many names so that my time goes toward the calls that need it, not the bookkeeping that drowns it.
The agents are not the analysts. The agents enable the analyst-equivalent function. The actual analyst — the one who reads the surfaced material, weighs the persona arguments, makes the call — is me, just with a much wider field of view and a much sharper set of frames than I ever could have maintained on my own.
If you want one principle that compresses everything I've learned in the last three months of building this, it's this: the agent team does not make the human obsolete; it makes the human the bottleneck where the human was always supposed to be. All the rest is tooling.
A request, to anyone reading this
If you are a professional allocator, a serious retail investor, or someone who has watched the markets and felt the gap between the work you wish you could do and the work you actually have time for: start playing with agentic AI tools, today, now. Not next quarter, not when the platforms are mature, not when someone else has shown the path. The compounding starts the day you begin, and the gap between people who start now and people who wait will, I suspect, look a lot like the gap between people who started spreadsheeting in 1985 and people who waited for it to be simpler.
The tooling is rough. The prompts will fail. The first cron will break midway. The first persona will sound generic. The first pipeline will mess up its outputs. None of that matters; the leverage is in the practice. You are not building software. You are grooming a team, your own team.
You will discover, as I did, that the work is not technical. It is editorial. Deciding what each persona should believe, what each pipeline should refuse to do, what each monitor should alert on — that is the job. The tools are commodities. The judgment is the moat.
Where this is going
I don't know what the next quarter or year of this looks like. The architecture I've sketched here is a snapshot, not a destination. Some pieces will be replaced. Some will be deprecated. Some will be rebuilt from the ground up when a better tool ships.
What I do know is that the direction is right. Solo investors with the discipline to build their own analyst team, and the humility to keep their own judgment at the centre of it, will produce work that genuinely couldn't have existed five years ago. That is the kind of difference that shows up in returns over a decade — not because the AI picked the stocks, but because the human who picked the stocks had a team that finally let him or her see the field clearly.
If anything in this is useful to you, take it and run. If you're building something similar, write about it — the field is too young to have a settled best practice, and the people who share what works will compound faster than the people who hoard. I'd rather lose the edge of secrecy than lose the contribution to the practice.
Onward.
Alpha Mason
Construct Alpha is a personal record of one solo PM's process. Nothing here is investment advice. Specific positions, current portfolio holdings, and live trade-level decisions are deliberately omitted from public writing.*