Most projects die as demos. We ship the ones that don't.
onebrain designs and builds custom platforms, data pipelines, ML models, agents and RAG systems that survive contact with production. Software architects and a forward deployed engineer, not a prompt wrapped around someone else's API.
Shipped, not pitched
Software that's already running.
We don't have a slide about our methodology. We have products in production, some with models in them, some without, and the people who built them are the people you'll be talking to.
Scan product
Retail operations
TuStockYa
LIVE
An operating system for sneaker retail. Photo recognition identifies a shoe in seconds, inventory stays true across every store, and the point of sale runs on a phone instead of a terminal.
- ScanPoint the camera at the product
- IdentifyModel and reference in seconds
- SellLive stock across every store
Built by Juan Sebastián Medina.
View TuStockYaNocturne Op. 9 No. 2
your piano sounds at the same time
Connected hardware
Forte
LIVE
Turning an acoustic piano into a connected instrument: the mechanics stay, the intelligence is new.
- Score + pianoEvery note lights its key, instantly
- Tutor + pianoWhat the teacher plays sounds on yours
Built by Xamir and Juan Sebastián Medina.
View Fortecore/ the rules it starts with memory/ what the business knows episodic/ what happened, dated semantic/ facts, clients, decisions procedural/ how things get done inbox.md raw capture, nothing lost ventures/ offers · pipeline · clients artifacts/ what actually gets read
Shared memory
The memory system
LIVE · in daily use
Rules, history and decisions in plain files that people and agents read the same way. No database, no vendor. It runs a whole operation every day.
Built by Jhonalbert Álvarez.
View the project› touch.draft
reason public record, SECOP
✓ Ley 2300 sending window
✓ evidence attached
✗ blocked invented price
› waiting for human approval…
Outbound with approval
Icarus
BUILT · v0
An agent that drafts every outbound touch with its reason attached, runs it through hard gates (Ley 2300 sending windows, evidence required, no invented price) and waits for a human to approve. Nothing sends itself.
Built by Jhonalbert Álvarez.
View the projectInventory questions
The inventory agent
BUILDING
For a Colombian pharmacy retailer. Answers inventory questions from the team's own spreadsheets, in Slack, for everyone. The model never sees the data and never writes a query: it picks a metric, the system runs it.
Built as a team
A sourced briefing
The personal assistant agent
BUILDING
For the same retailer. Reads a manager's mail, calendar, Slack and Drive twice a day and tells him what needs him, with the source of every item attached.
Built as a team
Part of the team comes from robotics and autonomous systems, where software that mostly works isn't a category.
What we do
The outcomes. Everything else is implementation detail.
- 01
Ship the product
From idea to production, not to a demo.
- Custom platforms
- Software development
- 02
Make it intelligent
Models that answer from your data, and cite where the answer came from.
- Real AI
- RAG
- ML models
- 03
Make the data usable
One source of truth instead of a pile of spreadsheets.
- ETL
- Data engineering
- 04
Make the work run itself
The busywork happens without anyone doing it.
- n8n
- Ops automation
Not sure which one you need? That's what consulting is for. It's the door, not another product. And sometimes the honest answer is that your problem doesn't need a model at all.
Make it intelligent
Real AI means it's still right on Tuesday.
A demo answers the question you rehearsed. A system answers the one your customer actually asks. From your documents, with a citation you can click, and a logged trace when it gets something wrong. We build retrieval that's evaluated, not vibed: a test set, a score, and a number that has to go up before anything ships.
A chat wrapper
Which contracts auto-renew in Q3?
Based on typical contract structures, many agreements include auto-renewal clauses that activate unless notice is given within a specified window…
- No sources
- Never read your documents
- Sounds right, says nothing
onebrain
Which contracts auto-renew in Q3?
Four do. Northwind renews first, on July 14.
- 1MSA_Northwind.pdf · p.12
- 2Acme_Logistics_amendment.pdf · p.3
- Every claim traceable
- Read your documents
- Not a thin wrapper around a chat API. Retrieval you own, models you can retrain, and pipelines that run where your data already lives
- RAG over your own documents, where every claim traces back to a source
- Evaluation sets and retrieval scores, so "better" is a measurement and not an opinion
- Custom models when a prompt genuinely isn't enough. Trained, versioned, and monitored (TensorFlow, PyTorch, MLflow)
- Guardrails and logged failures, because the useful question is what it does when it's wrong
Make the data usable
Every source. One answer.
Your data isn't missing, it's scattered. A CRM, a few spreadsheets, a Postgres nobody documented, and an API that rate-limits at the worst possible time. We build the pipelines that pull it together and keep them boring: idempotent, monitored, and cheap to re-run when something upstream changes.
Scroll to assemble
Ingest
Databases, APIs, files, and the spreadsheet someone maintains by hand.
Transform
Versioned, reviewable transformations. The logic lives in code, not in a formula nobody can find.
Orchestrate
Scheduled, idempotent runs. Re-running is safe, so recovering is boring.
Observe
When a source changes shape, you hear about it before your dashboard lies to you.
Make the work run itself
The report that writes itself before you ask.
n8n is where most of your operational tax disappears. A deal moves and the CRM updates. Monday arrives and the KPI brief is already in the shared folder. A document changes and the index that answers questions about it changes with it. We wire the systems you already pay for into each other. And we make the failures loud, because a silent automation is worse than no automation.
Monday 09:00
0.4s
Trigger
Northwind moves to "Closed Won"
Enrich
Pull the account history
Decide (AI)
Renewal risk: low → no escalation
Act
Update the CRM · file the contract
Notify
Brief in #revenue, folder updated
Nobody touched anything
If any step fails, you hear about it
- Workflows across the tools you already use. CRM, docs, Slack, billing, your database
- AI in the loop where judgment is needed, deterministic code where it isn't
- Alerting on failure, retries with backoff, and a log you can actually read
How we work
Ways to start.
Diagnostic
Most engagements start here
- Duration
- Short and fixed
- Scope
- Fixed
- Deliverable
- Architecture + build plan
- Demos
- Not included
- On-call
- Not included
Build
- Duration
- Per project
- Scope
- Defined
- Deliverable
- Product in production
- Demos
- Weekly
- On-call
- Not included
Embedded
- Duration
- Monthly
- Scope
- Open
- Deliverable
- Your architecture team
- Demos
- Weekly
- On-call
- Included
Whichever way you start, the architecture is yours to keep.
Scope drives price, so we quote once we understand the problem. That's what the call is for.
Who you're actually talking to
Machine learning to sales. The people who build it.
Xamir
Software Architect - robot.com
Software Architect at Robot.com. Robot.com, formerly Kiwibot, operates one of the world's largest autonomous robot fleets, from sidewalk delivery to warehouse logistics. He works across machine learning, backend, frontend and DevOps, in TensorFlow, PyTorch and MLflow. At onebrain he's the consultant, the one who tells you whether your problem needs a model or a better schema.
Juan Sebastián Medina
Software Architect
Software Architect of TuStockYa. TuStockYa is an operating system for sneaker retail: photo recognition identifies a shoe in seconds, inventory stays true across every store, and the point of sale runs on a phone instead of a terminal. He built it across machine learning, backend, frontend and infrastructure.
Jhonalbert Álvarez
Forward Deployed Engineer
He works across direct-response copy, delivery inside the client's operation, sales and GTM, RevOps, and data and AI, in Python, Postgres, Claude Code, MCP servers and skills, with local models for what cannot leave the machine. At onebrain he's the one on the call who turns your problem into a working agent inside the stack you already run.
Visit Jhonalbert Álvarez's site (opens in a new tab)No account managers, no offshore handoff. The people on the call are the people who build it, all of them shipping AI inside real companies, not slide decks.
Next step
Bring us the problem you can't stop thinking about.
A short technical call, no deck. Come with the messy version: the spreadsheet, the process nobody wants to own, the model that works in the notebook and nowhere else. You'll leave with an honest read on whether it's worth building and what it would take.