The self-improving system
Your firm’s 7 years of knowledge, finally compounding, while you sleep.
A knowledge base that reads across everything your firm knows (client history, playbooks, the judgment calls nobody wrote down) and connects it overnight. Built on Claude Code and plain files. Yours to own outright.
- It improves itself overnight. You don’t feed it. Every night it re-reads what you’ve added and finds connections you didn’t ask for. You wake up to what you’d have missed.
- Built on tools you can get today, not exotic tech. Claude Code and plain text files. Nothing that hasn’t shipped, nothing you’ve never heard of, nothing locked behind a vendor.
- You own the whole compounding asset: code, files, all of it. It appreciates as your firm learns. A SaaS subscription depreciates the day you stop paying. This is yours.
@you how did we price the SkyBridge job last year? Client’s asking again.
Re: The Hartwell workaround
Did we ever document how we solved this? Can’t find it anywhere.
Last edited 2023. Nobody owns this folder.
412 docs · nobody trusts them
Handover call · the partner on leave knew this, not you
“I know we solved this before. I just can’t remember where.”
Watch first
See it find a connection overnight.
Watch: how the system found a connection overnight that I’d have missed for months. (4 min)
I walk through my own setup on real data. No slides, no theory. If the overnight claim sounds like hype, this is where it stops sounding like hype.
The blueprint
Get the exact blueprint I run for my own firm. Free.
The real setup, not a teaser: how the system reads, connects, and surfaces overnight. The same one running behind this page.
Send me the blueprint →How it actually works
How it gets smarter without you touching it.
For a claim like “it compounds overnight,” the how is the proof. Here’s the whole mechanism: three steps, nothing hidden.
Reads across everything.
It ingests your client history, past projects, playbooks, notes, decisions: plain files it can read end to end, not folders it dumps and forgets.
Connects it overnight.
While you sleep, it re-reads the whole base and looks for links between things filed months and clients apart. New knowledge you add tonight gets connected to everything already there.
Surfaces what you’d have missed.
In the morning you get the connections worth acting on: the “this client’s problem is the one we solved for someone else two years ago” moment, handed to you instead of lost.
A concrete one
On 12 June 2026, a prospect asked me a pricing question I was sure I’d never handled. Overnight, my own system connected it to a scoping call from November 2025 (a different industry, a different client, but the same underlying constraint) and surfaced the exact number I’d landed on back then and why. I hadn’t tagged either. I hadn’t linked them. I’d have spent an afternoon re-deriving it, or worse, guessed. The system had already done the work while I slept. That’s the entire product: not a place to store what you know, but something that keeps finding what you already knew and forgot.
Everything you know, scattered.
In your head. In two other people’s heads. In an inbox nobody owns, a drive nobody trusts, a doc last touched in 2023.
Everything you know, connected, and compounding every night.
One base the whole firm can query. Every new piece linked to everything already there. The green node pulses because the connections keep forming while you’re not looking.
The mechanism, made visible
One base. Every night, the connections multiply.
This is what “compounding” looks like: everything your firm knows, pulled into one place, then re-read and re-linked every night into a single brain that gets smarter on its own.
What goes in
- Client history: every account, every past engagement, every thread
- Playbooks and SOPs: the way you actually do the work
- Past projects and proposals: what you scoped, priced, shipped
- Call notes and decisions: the judgment calls nobody wrote down
What comes back out
- Connections between things filed months and clients apart
- “We solved this before,” surfaced before you re-solve it
- A plain-language answer to anything the firm has ever known
- A record that stays, even when the person who knew it leaves
Sound familiar?
Where does everything your firm knows actually live?
Seven years in, the answer is almost always the same, and it’s the answer that keeps you up at night.
It’s in your head. You’re the search engine. Every “how did we handle X” routes through you, and it always will until you write it all down, which you never will.
It’s in one other person’s head. If they take two weeks off (or leave), a chunk of how your firm actually works walks out the door with them.
You tried Notion. It rotted. It became another folder nobody updates, because keeping a wiki alive is a job, and nobody has that job.
The same problem gets solved from scratch, twice. Something you cracked for a client in 2024 gets re-figured-out in 2026 because nobody remembered it existed.
Your best judgment never gets written down. The pricing calls, the “don’t take this kind of client,” the workaround that saved a project. The most valuable knowledge you have is the least documented.
The proof
The proof is that I run this myself, and it’s already saved my own firm work.
The blueprint you’re opting in for isn’t a concept. It’s the system behind this page. Below it: four builds I’ve shipped, each one turning trapped knowledge into something the firm keeps.
Lead exhibit · my own system
My own system.
It reads across every scoping call, every proposal, every build I’ve done since I started. It runs overnight and hands me connections in the morning. The 12 June example above is one of dozens. I didn’t build this to sell it. I built it because I was the bottleneck in my own firm, and I got tired of being the only place my knowledge lived. The blueprint is exactly this, documented so you can see how it’s put together.
Past searches, finally compounding.
A Singapore senior-placements firm did 8 hours of manual market mapping per candidate, much of it repeating work done six months earlier on a similar search. I built an automated sourcing pipeline. The pool grew 2–3×. And the part that compounds: every past search is now searchable, so the firm stops re-doing work it already did. The knowledge accumulates instead of evaporating.
−8h per candidate · 2–3× pool · custom build
A decade of statements, one live view. And it stays.
A Singapore family office held positions across multiple private banks, each statement a different PDF format, consolidated by hand every week. I built an OCR + consolidation + risk-flagging system. The director sees the full portfolio live, without asking anyone. Every statement the system parses stays parsed. The historical record builds itself instead of living in a stack of PDFs nobody can query.
multi-bank OCR · live portfolio view · custom build
Four retailers’ data, unified. And every month retained.
A Singapore last-mile 3PL ran performance analysis across four retailer accounts, each in a different Excel format on a different cadence. Days of manual compilation. I rebuilt the pipeline and dashboard: four feeds into one, year-on-year and SKU-level, answered in seconds. Every period that flows in is kept and comparable. The operating history compounds instead of resetting each month.
4 feeds → 1 · SKU-level live view · custom build
The knowledge left the weekly meeting and moved into the system.
A Singapore education consulting firm ran a 3-hour weekly sync just so everyone could update everyone: the only time anyone had a complete picture, and it was wrong by Wednesday. I unified the work into one platform. The 3-hour sync is gone. No new admin hire. Student count per consultant went from 12 to 18. What used to live in one meeting a week now lives in a system the whole firm can query any day.
3h weekly sync killed · 0 new hires · 12 → 18 per consultant
The honest comparison
The honest comparison.
Five ways to deal with what your firm knows. Four of them you’ve already tried.
| Comparing | Static wikiNotion / Confluence | Knowledge in headsYou and two other people | Generic ChatGPTCheap seats, no memory of you | A self-improving system you ownBuilt by Ben · yours outright |
|---|---|---|---|---|
| Year-one cost | Low subscription, plus the unpaid job of keeping it updated | “Free,” until someone leaves and takes it with them | Cheap seats, plus your data pasted into someone else’s model | One build. Then it’s yours, no subscription. |
| Gets smarter on its own? | No. It rots the day you stop feeding it | No. It only knows what today’s people remember | No. It forgets you the moment the chat closes | Yes. It re-reads and re-connects every night |
| If a key person leaves | You lose whatever they never wrote down | You lose a chunk of how the firm works | Nothing was ever captured to lose or keep | Their knowledge is already in the base, connected and searchable |
| Who owns it | The vendor. You rent access | Whoever’s head it’s in | The vendor. And they train on the industry, not on you | You. Code, files, the whole asset |
| What you’re left with | A dead folder nobody trusts | A bottleneck named you | Generic answers with no memory of your firm | A proprietary asset that appreciates as your firm learns |
If it hasn’t surfaced a connection you’d have missed within 30 days, I rebuild it, or that cycle is free.
The whole claim is that it finds things you’d have lost. If it doesn’t do that inside a month, I haven’t delivered, and I don’t get to keep pretending I did.
Questions
The things founders ask before they opt in.
Straight answers. If yours isn’t here, email ben@3nm.io.
Is my data safe?
Yes, and it never leaves your control. The system is built on plain files that sit in your own storage, on infrastructure you own. Nothing is pooled into a shared model, and nothing trains anyone else’s product on your firm’s knowledge. On delivery, the whole thing is yours to host wherever you want.
What if I’m not technical?
You don’t touch the code. I build it, hand it over documented, and it runs on its own overnight without you managing it. Using it is asking it questions in plain language. Notion asked you to become a librarian; this doesn’t.
Do I own it?
Completely. Every line of code, every file, the whole system becomes yours on delivery. No SaaS seat, no lock-in, no “cancel and lose it.” It’s an asset on your side of the table, and it appreciates as your firm adds to it.
How is this not just Notion?
Notion is a place you put things and then have to maintain. This reads across everything on its own and connects it overnight. The part where a human keeps it alive is exactly the part I removed. Notion stores. This one thinks about what you stored while you sleep.
What’s actually delivered?
The self-improving knowledge system built on your firm’s real data: the ingestion setup, the overnight connection engine, a plain-language way to query it, and the plain files underneath, all documented and handed over as yours.
How long does it take?
A working version on your real data early, then the full build inside a fixed window, same cadence as every build on my case-study page: scoped in writing within 48 hours, something real by day 7.
What does it cost?
No number on this page on purpose: the build scales with what your firm knows and how it’s stored, and I won’t quote you blind. We settle it on the call, in writing, before anything starts. Start with the free blueprint; the cost conversation only happens if the blueprint makes sense to you.
Who builds it?
I do. Ben: one operator, direct, no agency layer and no handoff to a junior. The system running behind this page is mine; the one I build for you is built the same way.
Get started
Start with the blueprint.
It’s the same system running behind this page.
Get the exact setup: how it reads, connects overnight, and surfaces what you’d have missed. Then decide if you want me to build yours.
Get the self-improving system blueprint →Not ready to build? Leave your email anyway. I send one note when I ship something worth your time, no drip machine, no chasing. When your firm’s knowledge finally becomes the thing you can’t afford to lose, you’ll know where to find me.