Hi, I'm José. Chemical Engineer & MSc in AI.
Chemical engineer, now doing an MSc in AI, and the builder of the agentic system this page describes. Several LLMs, one of them a model of my own that runs offline, share one memory of everything I know, follow written rules, log every change and learn from their own mistakes while a nightly routine checks their work. The same setup produces the engineering tools I design with, a handful of apps and a game. Every number on this page is measured, not estimated.
- 336,170lines of Python, written
- 3,213changes logged by agents
- 32,089nodes in its memory
- 187lessons learned
01Portfolio
Six builds and 224,884 of those lines of Python, each with a receipt.
02Agent operating system
Four agents, one workspace, fifty-six rules.
Claude, Codex, Gemini and my own local model share one agentic workspace I call Igni. Each claims a path before it edits and logs every consequential change to a shared ledger. A 46-step daily run backs up, audits and re-indexes the lot.
Agentic Office · drag to rotate · scroll to zoom
00Igni · how I think
Igni: how my mind works, written down.
Igni is the agentic architecture I work inside: a knowledge graph for what I know, written rules for how I work, and a harness that loads both into every LLM agent before it touches a file, whichever model it is, my own included. When an agent gets something wrong, the fix becomes a lesson; a lesson that repeats becomes a rule. Less an app than a way of working the agents can read.
- a mistake
- a lesson
- it repeats
- a rule
- every agent, next session
- Each file becomes nodes and the links between them. An agent asks a question in plain words and gets back the few files that answer it, with the path to open, instead of reading the whole tree.
- Rules say what never to do and why. Skills are step-by-step recipes for a kind of job. Hooks run on their own at the start and end of a session, so nothing depends on the agent remembering.
- A claim is a line in a shared file: who, which path, since when. A second agent that asks for the same path is refused. Every change ends with an entry in the ledger, so tomorrow’s session can read what today’s did.
- The harness routes each request to the right skill, gives the model its tools and checks the result before it counts. A capability isn’t done until it passes the same test gauntlet every time.
- One script runs every step in order and writes a one-line status for each. A checker that can’t run is reported as broken, not as clean, and the morning brief says what needs me.
- Each lesson records what happened, why and what to do instead. Every session opens with the latest ones, so a mistake made on Monday by one model isn’t repeated on Tuesday by another.
- It reads the same status files the agents write. Green checks make it calm; a failed step or a stale claim changes its expression, so I see trouble before I read a log.
Igni · drag to rotate · scroll to zoom
03Retrieval
Every question starts in the brain.
A neural link to everything I know: projects, files, notes, drawings and code in one knowledge graph, wired into the whole system and served to the agents as a local MCP server. Query it before opening a file; exhaust it before calling anything missing.
Drag to rotate · scroll to zoom
- 32,089 nodes
- 50,717 edges
- 3 MCP tools: search, explain, path
04Engineering tools
Ten tools that draw and check what a plant needs.
Most are skills: written procedures with their checks built in, so nothing ships until it passes. Open any of them for the detail.
05Foundations
From LNG terminals to water plants, now building with AI.
Experience
- 2026–Project EngineerWater-treatment engineering · IrelandFiltration, softening, RO, DI, UF, chemical dosing and UV: design, specification, suppliers and commissioning.
- Technical point of contact for pharmaceutical, food, manufacturing and municipal clients, from enquiry to commissioning.
- Brought AI tools into the team’s daily work: document search, drawing and proposal generation, automated quality checks.
- 2023–25Operations ManagerFood distribution · USAOperations, the purchasing and contracts team, sales targets, import-cost negotiation and inventory KPIs.
- Accountable for monthly commercial targets and client retention.
- Resolved customer and supplier disputes on quality, quantity and delivery times.
- 2023Technical ManagerFlotation technology for mining · ChileLed engineering, lab, projects and admin; restructured a team of 13 and brought the technology to large miners in Chile and Canada.
- Client-side pilot-plant campaigns that proved the technology on the client’s own site.
- Introduced performance metrics, standardised functions and redefined roles; reported to the US parent company.
- 2021–23Product ManagerElectronics import · ChileLogistics, sales and customs; the product lifecycle, the annual sales plan and supplier negotiation in China.
- Ran the full import and customs-clearance process.
- Prepared the annual sales plan and monthly reports.
- 2019–21Head of Process Engineering & Energy EfficiencyLNG regasification terminal · ChileEnergy-efficiency and pre-feasibility studies, green-hydrogen and energy-recovery screening, ISO 50001 certification, commissioning of ultrasonic meters and cryogenic pumps.
- Chaired interdisciplinary committees (process, mechanical, electrical, instrumentation, operations) that approved plant changes on a live, safety-critical site.
- Member of the technical committee of the national natural-gas industry association.
- 2015–19Process Engineering & Metering SupervisorLNG regasification terminal · ChileProcess engineering and the gas-quality lab, national consumption forecasts, CAPEX/OPEX, ISO 6974/6976 metrology, ISO 9001/14001/OHSAS 18001 certification.
- Monthly and annual reporting to the board.
- Maintenance plans for the terminal’s analytical and metering instruments.
Education
Certifications
- ISO 50001 Lead Auditor (IRCA)
- Aspen HYSYS dynamic modelling
- ISO 9001 · 14001 · 18001 · 27001 integrated systems
- ISO 17025 laboratories
- Ultrasonic meters & gas chromatographs
- Metrology & measurement uncertainty
- Gas metering systems & stations
- Tank-level instrumentation
- Process gas chromatograph O&M
- Flotation pilot & demo plants (USA, Canada)
Community Claude Community, Cork · AIS26, Dublin
Languages Spanish, English
Software Python · JavaScript · MATLAB · VBA · Git · Electron/Node · Aspen HYSYS · PIPE-FLO · AutoCAD · OSIsoft PI · Power BI · SAP · TIA Portal · Blender · Unreal Engine 5
AI Claude Code · Codex · Antigravity · Ollama · MLX · MCP · RAG · knowledge graphs · eval harnesses · LoRA
06Own model
My own model, running offline.
An open model on my laptop inside Raszagal, a harness I wrote. Given a written brief, it picks my process-diagram skill and returns the finished CAD drawing on its own: 119 s the first time.
Raszagal routes each request to a skill, retrieves from my knowledge graph, and verifies answers by running them instead of judging them: code pass rate went from 72 to 89 % on that change alone. Anything that came in through web or mail is tainted and cannot run code until I approve it.
- Base
- Qwen3.6-35B-A3B, open (Apache-2.0)
- Experts
- 256, 8 per token, ~3B active
- Size
- 4-bit, 23 GB, 256K context
- Runs on
- one laptop, fully offline; speaks the OpenAI API
- Speed
- 47–64 tokens/s measured locally; +46 % on the MLX runner
- Harness
- 28,168 lines, 16 tools, 19 skills
- Safety
- web or mail input locks code until I approve
- Tested
- 163 cases, 8 axes, no AI judge
- Training
- LoRA on my own architecture.
Zeratul · drag to rotate · scroll to zoom
Gauntlet by axis17-Sep-2026 · 120 cases · three seeds pooled · pass rate
Why this base12-Jul-2026 bake-off · same 44-case suite · deterministic checks, no AI judge
07Lab
Where new tools are tried before they earn a place.
Where new tools are tested before they join a new build programme version: seven leading open-source repos for water and process engineering on GitHub, which I am rebuilding as new, upgraded versions.
08Inventions
Six apps, one system.
Six apps built around the system: one decides what the agents may do, one is its face, one types what I say, one listens to my music, one pins a window on top and one lives in my pocket.
Reviews every Claude Code and Codex permission request from the menu bar: one plain sentence on what will happen, a risk level, and more friction for riskier requests.
Swift · menu barThe system's own face: its expression follows the measured state of the workspace, and it answers questions from the knowledge graph.
Electron · cockpitHold a key, speak, let go: the words land at the cursor in any app, transcribed and tidied on my own Mac, nothing sent anywhere.
Swift · local WhisperA Logic Pro companion that listens to the project: tempo, key, groove, drum grid, chords, loudness, and the closest sound in my own sample library.
Electron · Python enginePins any window on top of the rest. An open-source app I compiled from source, patched and signed on my own Mac so it keeps working without its update feed.
Swift · built from sourceA live AI companion for the iPhone: screen always on, a wake word, streamed speech you can interrupt, and a model that runs on the phone itself.
Swift · iPhone · on-device09Off hours
Where the same tools go to play.
The Hammer Falls marine · walk cycle · drag to rotate · scroll to zoom · Marine and C-14 rifle model by dwelfster, CC BY
↵Let's talk
Hiring, collaborating, or curious how my mind works?
Find me on these.
›Search