What needed solving
In rehabilitation, a small variation can be progress or just camera noise. Without knowing that error floor, the number becomes more convincing than true.
Seven works on show.
From rehabilitation to the daily operation of a company. Scroll and the room moves; open any of them to read its full dossier.
No. 01
KINELAB
Rehabilitation measured with the browser camera; every metric knows its own noise.
No. 02
Nexo
Meals and glucose on the iPhone, local-first; the AI analyzes off the phone.
No. 05
Agent Mini Deck
A local command center for Claude Code, Codex, OpenCode and Cursor — on the Mac, the iPhone and a Stream Deck.
No. 07
Segundo Cerebro
The day in audio → transcription → analysis → Obsidian. A cognitive mirror.
No. 13
GPS 3.0
The commercial and logistics operation of a real company, in production.
No. 18
SISMO
Earthquakes, weather and 116 of Chile's critical assets crossed in near real time.
No. 19
ArchiveAll 19 works, one by oneFRAUDE
Fraud campaigns in Chile detected and correlated from public sources.
All the works · 01—19
§1 19 systems. One practice.
Health, personal AI, work and family. Explore each project: what needed solving, how it works and where its limits are.
Lens
HealthMeasuring without inventing certainty. Tools that explain their noise and their limits.
01KINELABRehabilitation measured with the browser camera; every metric knows its own noise.HealthPrototype
The piece
It measures joint angles, counts repetitions and tracks progress using only the camera. A change is only asserted when it exceeds what noise can explain (MDC95, Theil–Sen). The video never leaves the device.
Tools
JavaScript · MediaPipe
Principles applied
Local-first · Known noise
The number
MDC95 calculado por usuario
How it works
MediaPipe detects 33 landmarks on the device; filters, personal calibration, MDC95 and a Theil–Sen trend turn the signal into angles, repetitions and defensible changes.
What it refuses to do
It is an instrument of trend and adherence, not a certified goniometer or a diagnosis. The video never leaves the device.
02NexoMeals and glucose on the iPhone, local-first; the AI analyzes off the phone.HealthIn progress
The piece
No account, no cloud of its own: everything lives in a versioned JSON on the device. AI analysis runs on a Mac through an iCloud folder. Clinical guardrails: never doses, never future glucose.
Tools
SwiftUI · Swift 6
Principles applied
Local-first · Human gate
The number
~270 tests unitarios
What needed solving
An isolated glucose reading doesn't explain what happened. Understanding it means joining meals, measurements and context — without turning health data into yet another cloud account.
How it works
SwiftUI keeps a versioned JSON on the iPhone; an iCloud folder connects to the analysis on the Mac, and a suite of ~270 tests protects the clinical and sync rules.
What it refuses to do
It doesn't compute doses, doesn't predict future glucose and requires no account. The AI analyzes off the phone, but makes no clinical decisions.
03
GARMINA personal health panel that ends in a PDF to bring to the doctor.HealthMature
The piece
Heart rate, HRV, sleep and glucose consolidated into reports a doctor can read. Alerts are deterministic; the model only writes.
Tools
Python · pandas
Principles applied
Known noise · Rules + AI
The number
4 señales → 1 informe médico
What needed solving
Heart rate, HRV, sleep and glucose arrive in separate formats and don't form, by themselves, a story a doctor can review quickly.
How it works
Python and pandas consolidate the four signals, apply verifiable rules and produce a readable PDF; the model only helps write the summary.
What it refuses to do
It neither diagnoses nor invents alerts. The public image is limited to an anonymous statistical output — no names, dates or individual clinical values.
04
NUTRICOA label scanner for Chile's food-warning law (20.606).HealthSeed
The piece
OCR plus verifiable rules for classification; the LLM only suggests alternatives.
Tools
Expo · Open Food Facts
Principles applied
Rules + AI
The number
Ley 20.606 convertida en reglas
What needed solving
Reading a label and relating it to the Chilean warning-seal law takes more than recognizing text: the thresholds must be reproducible and explainable.
How it works
OCR captures the data, rules encode Law 20.606 and Open Food Facts completes the record. AI comes in afterwards, only to suggest alternatives.
What it refuses to do
It is at seed stage. The legal classification is not delegated to a model, and no suggestion replaces professional advice.
Agent operationsLocal control, evidence and continuity for working with several agents without losing the human gate.
05Agent Mini DeckA local command center for Claude Code, Codex, OpenCode and Cursor — on the Mac, the iPhone and a Stream Deck.Agent operationsIn production
The piece
It turns agent work into an operable surface: it shows what every session is doing, keeps context across devices and makes visible the points where a person must step in. Not another chat — the control-and-evidence layer that joins four agents with three physical surfaces.
Tools
Node.js · HID · PWA · Tailscale
Principles applied
Local-first · Human gate · Rules + AI · Build
The number
4 agentes · 3 superficies
What needed solving
When several agents work at once, their state, questions, permissions, plans and evidence end up scattered across terminals. The operator loses continuity exactly when deciding matters most.
How it works
Hooks from Claude Code, Codex, OpenCode and Cursor feed one shared state. The same daemon turns it into Stream Deck controls, a private iPhone PWA and the deck CLI — with sessions, races, handoffs and observable verification.
What it refuses to do
Mini Deck doesn't self-authorize destructive actions and never confuses an answer with verified work. The important gates still demand evidence and an explicit human decision.
06
PROJECT ATLASClaude, Codex and Grok deliberate blind on a shared board; every decision is ratified by the human.Agent operationsActive
The piece
Not another system of agents chatting: a measuring instrument. Three model families answer blind, critique each other by alias, and a human ratifies with the dissent in view; every approval is a gold label that says when deliberating paid off and when a single model would have been enough.
Tools
Python · JSON Schema · Claude Code · Codex CLI · Grok
Principles applied
Known noise · Human gate · Rules + AI · Build
The number
3 modelos · máx. 2 rondas
What needed solving
Asking several models the same thing produces parallel chats nobody audits: there is no way to know when multi-model deliberation beats a single model, and no record of what was approved, why, and with what dissent.
How it works
One Markdown board per cycle, written only by a mechanical orchestrator; agents drop turns validated with JSON Schema. A Delphi protocol: a blind round, alias-based critique and a rotating critic issuing pairwise verdicts. Every cycle runs a single-model baseline and feeds a ledger anchored to the director's approval.
What it refuses to do
No agent executes or builds anything without the director's explicit approval; "done" is always a human verdict. The ledger still has too few cycles to separate signal from noise, and the real boards of each campaign are not published.
Personal AIListening, remembering and distilling without giving up control of the data.
07Segundo CerebroThe day in audio → transcription → analysis → Obsidian. A cognitive mirror.Personal AIActive
The piece
A lapel microphone and a pipeline that transcribes the day, identifies who spoke, classifies and pours it into a vault. The ambition: not just what I say, but how I say it.
Tools
Whisper · pyannote · Obsidian
Principles applied
Local-first
The number
1 día de audio → 1 registro
What needed solving
Everyday conversations and decisions get lost, and a chronological memory isn't enough to recognize patterns in what one says and how one says it.
How it works
Lapel audio, Whisper, speaker separation and classification turn the day into a structured record inside Obsidian.
What it refuses to do
The audio stays local, and the system works as a cognitive mirror — not as a record of others to publish, nor as automatic truth about a person.
08
TRANSCRIPTORA meeting copilot — live transcription and questions about what was said.Personal AIMature
The piece
Live ASR on the Mac, with model routing: the mechanical to local models, the hard to large ones.
Tools
MLX Whisper · NeMo
Principles applied
Local-first · Rules + AI
The number
2 micrófonos · audio local
What needed solving
In a long meeting you lose names, agreements and the exact minute backing a decision. A flat transcript doesn't make verification easy either.
How it works
Silero segments by utterance, Whisper transcribes locally with a glossary that learns, and two microphones can separate speakers by channel; every conclusion links back to its temporal evidence.
What it refuses to do
Recording requires informed consent. Audio is processed locally, and any query to an AI provider is an explicit choice.
09
HERMESA four-channel voice copilot, running locally.Personal AIIn production
The piece
Voice in, voice out, with an LLM brain — in daily production.
Tools
Python · Whisper · TTS
Principles applied
Local-first
The number
4 canales de uso
What needed solving
While driving, reading news or typing queries breaks attention. The information must also remain available when the connection fails.
How it works
A daily briefing and three conversational channels combine fresh sources, Whisper and local voice with Claude or Codex; Qwen stands by for offline stretches.
What it refuses to do
The channels don't execute arbitrary commands or write to the system. The site shows only the architecture — never real conversations or credentials.
10
YOUTUBEDiscovers, transcribes and distills AI videos in a single pass.Personal AIActive
The piece
A curation pipeline: from dozens of videos a week to a distillate of applicable techniques.
Tools
yt-dlp · Python
Principles applied
Build
The number
12 videos → 1 informe diario
What needed solving
Dozens of AI videos compete for attention, but few justify a full hour, and many claims come with no clear route to evidence.
How it works
The pipeline discovers, transcribes, classifies and analyzes; it separates short pieces from deep dives and builds a daily report with verifiable timestamps.
What it refuses to do
A timestamp proves the video said something — not that the claim is true. The report explicitly keeps that difference.
11
MargenAn academic brain on the iPad — read, annotate by hand, index everything and ask the AI that lives on the Mac.Personal AIActive
The piece
The place where studying happens and where what was learned is found again: the folder is the unit of study, everything produced gets indexed, and searching means being able to narrow down. Not a PDF reader with extras — the iPad's academic brain, with an AI that proposes and a person who decides.
Tools
Swift · PDFKit · PencilKit · SQLite FTS5 · Tailscale
Principles applied
Local-first · Human gate · Build
The number
5 formatos · 3 motores de IA
What needed solving
Studying on an iPad leaves the material as a pile of PDFs: notes, highlights and handwriting can't be found again months later. A note you can't retrieve is a note you never wrote.
How it works
A native iPad app for PDF, Markdown, HTML and blank notes; Apple Pencil ink, highlights, notes and tutor answers all enter the same SQLite index, with faceted and semantic search. It adds spaced review, mirror backups and a bridge to Obsidian; the AI runs on the Mac over Tailscale.
What it refuses to do
Reading, annotating, organizing and searching work offline; only the tutor needs the Mac. Codex and Claude credentials never leave the Mac, the AI proposes without applying anything on its own, and the site shows no real documents or notes.
12
Visor de TempsA native macOS app that relates temperatures, fans, memory and Ollama models to measure local AI.Personal AIMature
The piece
One concrete question — does an earlier fan curve yield more tokens per second? — and a tool built to answer it with data instead of intuition. It measures, compares and touches nothing: the decision about the curve stays with the person.
Tools
Swift · SwiftUI · Swift Charts · SMC/IOHID · Ollama
Principles applied
Local-first · Known noise · Build
The number
solo lectura · historial 5–30 min
What needed solving
When running local models on the Mac, you don't know whether heat is throttling the machine or whether an earlier fan curve would improve sustained performance. The data exists, but scattered across sensors, memory and Ollama.
How it works
A native macOS app that reads the SMC and IOHID sensors, asks Ollama which models are loaded and uses the system's official thermal state as the throttling signal. It keeps a 5-to-30-minute in-memory history and allows an A/B test with a fixed prompt, only when a button is pressed.
What it refuses to do
It is read-only: it writes no RPM or curves, loads no models and sends no prompts on its own; the curve is changed by hand in Macs Fan Control. It uses APIs Apple doesn't document, so it runs without a sandbox and isn't meant for the App Store.
Work · De Reyes AlimentosSystems that sustain the daily operation of a real company.
13GPS 3.0The commercial and logistics operation of a real company, in production.WorkIn production
The piece
Tracking, route planning, driver and salesperson portals, collections and ERP for a food processing company. A state machine with named human gates: authorizing and registering are decisions made by people with permission.
Tools
React · Express · PostgreSQL
Principles applied
Human gate
The number
+1.500 commits
What needed solving
Sales, routes, drivers, collections and inventory become isolated systems as a company grows; the cost shows up as double entries and context-free decisions.
How it works
React, Express and PostgreSQL sustain a multi-tenant platform with tracking, planning, field portals, collections and ERP modules connected through explicit states.
What it refuses to do
Authorizing and registering remain decisions of people with permission. This personal site publishes no production screens while they contain operational data.
14
dereyes.clA complete ecommerce — payments, electronic invoicing and stock synced from the warehouse.WorkIn production
The piece
A public store plus wholesale credit customers. Stock is proposed from the warehouse every morning; a person confirms.
Tools
Next.js · Webpay · SII
Principles applied
Human gate
The number
+18 años de operación digitalizada
What needed solving
A store for consumers and wholesalers needs payments, credit, electronic invoicing and warehouse stock that agree with each other, not a storefront detached from the operation.
How it works
Next.js connects the catalog with Webpay, Khipu and Chile's tax service; a daily process proposes stock from the warehouse before a person confirms the update.
What it refuses to do
The published capture comes only from the public homepage. Credit, invoicing and internal operations stay behind their access controls.
15
Agente documentalA photographed invoice becomes importable data — with a human gate.WorkIn production
The piece
It classifies its own output into four states; the value that cannot be wrong (the invoice number) is always set by a person. Only text leaves the machine.
Tools
Python · OCR
Principles applied
Human gate · Rules + AI
The number
4 estados de salida
What needed solving
Typing a photographed invoice into a management system is slow, but automating a wrong value can be worse than keeping manual entry.
How it works
OCR extracts the text and the agent classifies its own output into four states before proposing importable data for review.
What it refuses to do
The invoice number is always confirmed by a person, and only text may leave the machine. A real invoice is never used as the project's public image.
FamilyGames and tools made for specific people, not an abstract market.
16
ENTRENADORSpeed and angle of my son's shots, with computer vision.FamilyMature
The piece
A homemade sports coach: measurable metrics for every shot, to train by playing.
Tools
YOLO · MediaPipe
Principles applied
Known noise · Build
The number
velocidad + ángulo por tiro
What needed solving
To improve a shot it isn't enough to say it looked better. It needed measurable feedback that still felt like play, not like a laboratory.
How it works
YOLO and MediaPipe watch the trajectory and turn every attempt into speed and angle to compare training sessions.
What it refuses to do
This site publishes no names or identifiable faces. The chosen frame shows the session from behind, keeping the sporting context without exposing the child.
17
Isla de los SabiosAn educational game made to measure for my kids.FamilyDone
The piece
Learning by playing, with specs designed for two specific children — not for a market.
Tools
HTML5 · Canvas
Principles applied
Build
The number
3 misiones diarias
What needed solving
A generic educational game teaches an abstract audience; this one had to answer the interests, rhythms and challenges of two specific children.
How it works
HTML5 and Canvas bring together daily missions, memory, arithmetic and minigames like Comet Race in a tailor-made experience.
What it refuses to do
It is a family project, not a school platform. Public captures exclude names, faces and any data about the children.
Public data · ChileWatching the territory through open sources: every datum declares its origin and freshness.
18SISMOEarthquakes, weather and 116 of Chile's critical assets crossed in near real time.Public dataActive
The piece
Monitoring Chile's critical assets: CSN and USGS earthquakes, forecasts and measured air, crossed against mines, tailings dams, ports and hospitals. Risk decisions are documented, not hidden in the pipeline — and when two agencies disagree, it alerts on the conservative side of both.
Tools
Node.js sin dependencias · MapLibre · 20+ APIs públicas
Principles applied
Known noise · Rules + AI · Build
The number
116 activos · fuentes cada 5 min
What needed solving
Chile shakes every day, but knowing whether a quake matters means crossing it with what sits above: tailings dams, ports, 500 kV lines, reservoirs, hospitals. That crossing existed nowhere in one open place.
How it works
Every 5 minutes it crosses CSN and USGS — deduplicated, keeping inter-agency disagreement as uncertainty information — against a register of 116 assets, estimates intensity per asset and adds weather, measured air quality, swells and authority alerts from 15+ public sources. Zero npm dependencies.
What it refuses to do
It doesn't predict earthquakes and doesn't replace the authority (SENAPRED/SHOA is the official voice); it estimates intensities with public formulas and declares the source and freshness of every datum it shows.
19FRAUDEFraud campaigns in Chile detected and correlated from public sources.Public dataIn production
The piece
Homemade fraud intelligence: the central entity is the campaign, and every datum enters as an observation with its provenance. Strong matches propose relations; confirming is always a human act.
Tools
Python · SQLite · launchd · STIX
Principles applied
Human gate · Rules + AI · Build
The number
+1.100 campañas · ciclo cada 15 min
What needed solving
Fraud campaigns span domains, apps, social networks, phones and emails, but the signals live scattered across NIC, CSIRT, CMF and the platforms. Nobody joins them into a single case.
How it works
It censuses .cl domains, ingests authority alerts (CSIRT, CMF), watches social networks and correlates everything into campaigns where each observation carries source, date, confidence and provenance. It runs by itself in 15-minute cycles.
What it refuses to do
It never merges cases or reports accounts on its own: it proposes candidate relations and a person decides. The LLM describes; deterministic signals and the analyst rule. The panel is private — it shows active investigations.
§2 In numbers
+1,500commits in the largest system
~30automations in operation
~270tests in the glucose app
0personal data in the cloud
§3 Now
August 2026 — building Nexo, tuning the Segundo Cerebro, and serving this site from a computer at my home: no third-party cloud, no cost, behind a tunnel.
Local time in Concón
§4 Contact
This dossier remains open. If something you saw resonates — a project, a question, an idea worth building — write to me. I answer in Spanish or English.