A tailor made AI investment research system that compounds.
A Claude Code + Obsidian Notes research system that gives your team compounding investment alpha.
The problem · 01
Every session, traditional AI models start from zero. Your book doesn't.
Lost context, lost edge
Stateless models forget every session. The analyst re-explains the book every morning.
Conviction drift, undetected
The thesis quietly changes. Nobody logs it, and the position size no longer matches the argument.
Blind spots at scale
Cross-thesis connections vanish at forty names when positions are tracked in isolation.
Research that never compounds
Last quarter's work is stranded in a document nobody reopens. The book forgets faster than the people.
Vendor lock-in at the research layer
“AI for finance” tools built by people who have never run a book, holding your research on their servers.
The gap · 02
Everyone has the model. Almost nobody has the system.
Nearly every manager now uses a model. Almost all of that use is a chat window: summarise this, draft that, explain this filing. It saves reading time. It does not remember what was read, connect it to the positions held, or notice when a new datapoint contradicts an old claim. Allocators have started asking about the difference.
95%
of alternative managers use generative AI, up from 86% in 2023
AIMA, September 2025, 150 managers
2%
of AI use cases in UK finance run without a human. 34% of firms say they fully understand the AI they use.
Bank of England and FCA, November 2024
60%
of large allocators are more likely to allocate to a manager investing meaningfully in generative AI. 29% already ask about it in due diligence; another 29% will within a year.
AIMA, September 2025, 18 allocators
The fix · 03
Six things a chat window cannot do for a book of forty names.
| The problem | Today | With the system |
|---|---|---|
| The firehose | Thousands of items a day. A handful matter. Finding them costs the morning. | Scored, clustered and tied to the positions held before the day starts. |
| Reading is not filing | Insight stays in the reader's head. Cross-references decay within weeks. | Every source becomes a structured note, filed where it changes a claim. |
| Blind spots across positions | Nobody holds forty theses in their head. Shared assumptions go unexamined. | The whole book's research can be searched, and compared against itself. |
| Falsifiers never written down | Positions defended out of loyalty two years in. | Raise, cut and close conditions written in advance and tested on every datapoint. |
| Hierarchy compresses evidence | Analyst to sector head to PM: three hops and several weeks. | Evidence, thesis and falsifier live in one place, read whole. |
| Nothing compounds | Context rebuilt every morning, discarded every night. | Knowledge written back and cited months later. |
The system · 04
Laniakea Brain.
A research operating system on Claude Code and Obsidian. The whole team writes into it; the whole book lives in it.
The vault
A local, plain-text markdown knowledge base: theses, notes, sources, memos. Owned outright, version-controlled, no cloud lock-in, never used to train anyone's model. About 1,500 files on our own book.
The engine
Claude Code reads and writes the vault, runs deterministic skills, compiles research artefacts and records provenance. The analyst directs; it does the mechanical compilation.
Conviction that compounds
Persistent memory across every position, instead of resetting each session.
Six engines
Three interrogate the book (retrospective, cross-thesis, adversarial). One audits conviction against pre-registered triggers. One ingests the outside world. One compiles and publishes.
Judgement into procedure
27 skills and 8 portfolio-wide workflows, written around your process, so the same command gives the same shape of output every time.
Auditable by default
Runs on the firm's own machines. Every claim traces to a source. Answerable when an LP or an IC asks “why did we own this”.
The engine that ingests the outside world is fed by a sensing layer that runs while the desk is asleep. Here is one morning.
A morning · 05
How a morning runs.
The sensing layer runs on a schedule while the desk is asleep. It reads widely, writes only into a scanning folder, and puts the few stories that matter on your phone before the day starts. Nothing enters the book unless a person picks it.
- 07:00News sweep. About 6,800 items from five channels: hand-picked outlet feeds, Google News, an open news database, a paid news search and a financial-markets feed. Duplicates removed, headlines scored for new information, the same story from several outlets folded into one, full text fetched and re-scored. You see a brief grouped by theme and ticker, and the top stories on your phone.
- 07:30Catalyst reminders. Every dated event in the book, alerted two days out and on the day. A stale calendar flags itself.
- 07:35Price tripwires. Each thesis's raise, cut and close levels checked against the close. An alert quotes the trigger it threatens.
- 08:30X harvest. Posts on your names and themes tracked over days. The crowd's view read against your own bull and bear cases. Where the crowd makes an argument your thesis does not carry, you hear about it.
- ThenYou. Pick the one to three stories worth the book's attention. One command turns each into a research note and updates every thesis, sector and macro note it touches, with a dated log entry linking back to the source.
One morning brief, 15 August 2026
fetched
new since the last run
admitted after scoring
briefed
ingested by hand
Running cost $20–35 a month. Software free and self-hosted. Maintenance about half an hour a month.
The choice · 06
Buy, build or assemble.
Most firms will do all three. The question is which one holds the memory, because that is the one that compounds.
| Buy | Build | Assemble | |
|---|---|---|---|
| What it is | Vendor research platforms and the AI features inside the terminal | Machine-learning teams, proprietary models, cloud compute | Frontier models by subscription, a plain-text vault you own, procedure written once as templates and skills |
| What you get | Reading time back. Fast, governed, per seat. Identical to what every other buyer gets, and the memory it builds is the vendor's. | Differentiated signals, at tens of millions a year for the multi-strategy platforms. Long payback. | A compounding asset the firm owns, portable across models. The overnight layer costs $20–35 a month. |
| Who it suits | Firms that want a vendor to hold accountable | Large quant and multi-strategy platforms | Everyone from one analyst to a fifty-thesis book |
Laniakea Brain is the third column, built on your machines and written around your process.
Getting started · 07
What it costs to run, and how to start.
To run
- Obsidian: free.
- Claude: a subscription per seat.
- Sensing layer: $20–35 a month in model and search calls. The software is free and self-hosted.
- Maintenance: about half an hour a month, mostly pruning the watch list.
- Everything is plain text. If we disappeared tomorrow, the vault would still open.
To start
We start with a strategy sprint. We sit with the desk, map how research runs today, and write down where a system would change the outcome and where it would not. You keep the written roadmap. If it makes sense, we then build the system on your machines and train the team to run it.