Celaya Solutions Research
Start a build
CELAYA SOLUTIONS RESEARCHHIRE / 01

// 001 · Engagement

Twelve years on the floor. Now pointed at the model.

I am Christopher Celaya. I spent twelve years keeping industrial and critical infrastructure running: 38 kV gas-insulated switchgear at Schneider Electric, hyperscale critical facilities at T5 Data Centers, HMI and SCADA on a continuous manufacturing line at CN Wire in Santa Teresa. Now I build applied intelligence systems for the same kind of operation, and for the small businesses around them.

That order matters. Most people building these systems learned the software first and the plant floor never. It shows in what they ship.

Start with an audit hello@celayasolutions.com Ver en español
Based
El Paso, Texas
Serves
El Paso · Santa Teresa · Ciudad Juárez · Las Cruces
Languages
English and Spanish
Certified
NFPA 70E · OSHA 10

// 002 · Why this combination is rare

Two skill sets that almost never sit in one person.

There are plenty of engineers who can fine-tune a model. There are plenty of technicians who can trace a fault to a terminal. The number of people who have done both, in this region, in both languages, is very small, and every useful thing on this site comes out of the overlap.

What the industrial side taught

  • Assume failure. Redundancy exists because equipment breaks, and software that assumes a happy path is software that has never been on shift at 03:00.
  • Keep a record. A work order that cannot be backdated is worth more than a dashboard nobody can audit.
  • Never let an automated process act without a trace back to evidence and a path for human review.
  • A hardwired interlock outranks the controller. Always. In hardware.
  • The person who has to fix it at 3am gets a say in the design.

What that produces in the software

  • Local-first by default, because production data has rules about where it may live.
  • Citations on every claim, because a wrong torque spec has physical consequences.
  • Silence as a valid output. A system that says the corpus does not cover your question is more useful than one that invents an answer.
  • The language model stays out of the control path. It writes text, it does not move a breaker.
  • Measured results with the misses left in, not a benchmark chosen after the fact.

The longer version of that story

// 003 · What I build

Three kinds of work, one method.

Track 01

Manufacturing and industrial

For plants in El Paso, Santa Teresa and Ciudad Juárez

  • Downtime and fault prediction. Telemetry off the line into a model that gives hours of warning, so an unplanned stop becomes a scheduled repair.
  • Overall equipment effectiveness that people trust. Availability, performance and quality computed from measured data rather than from what got typed into a spreadsheet at end of shift.
  • Vendor documentation retrieval. Schneider, Rockwell, Siemens and your own drawings in one revision-aware search that cites the page.
  • Bilingual work instructions and standard operating procedures. Written once, answerable in the language the operator reads.
  • Visual inspection support. Defect classes a tired human eye misses at end of shift, with the disagreements routed to a person rather than auto-accepted.
  • Data capture from PLC, SCADA and historian. Usually the first job, because most of the value is trapped where nothing can query it.
See WITNESS, the industrial agent

Track 02

Small and mid-sized business

For the companies around those plants

  • Answering and booking systems. The after-hours call that currently goes to voicemail, turned into a job on the calendar.
  • Lead capture and routing. Qualified, tagged and in front of the right person before it goes cold.
  • Being findable by assistants. The buying question is moving from a search bar to an assistant, and most local businesses are invisible there.
  • Internal operations. Quoting, scheduling, invoice chasing and the paperwork that eats a Saturday.
  • Document intelligence. Three years of contracts, certs and inspection reports, answerable in a sentence.
The full service index

Track 03

Teaching and capability building

For teams who want their own people capable

  • Handover as a deliverable. Every build ends with your team able to run and change it. Dependency is not a business model I want.
  • Workshops for maintenance, quality and operations groups. What these systems can and cannot do, in plain terms, in either language.
  • Evaluation and honest scoping. Frequently the finding is that a proposed project is not worth doing, and saying so early is cheaper for everyone.
  • Curriculum work. I have built learning material for regional programs, including a curriculum for the El Paso Innovation Club.
How an engagement runs

// 004 · The region

This work is specific to the borderplex, on purpose.

A consultant flying in from Austin does not know that your maintenance techs read Spanish and your vendor manuals are in English, that your telemetry cannot legally leave the country it was generated in, or that the plant across the fence is running the same gear you are. I live here and I have worked in these buildings.

Ciudad Juárez

The maquiladora sector shed roughly 57,500 jobs between mid-2023 and mid-2025, about 18 percent of its manufacturing workforce, while output targets held. That gap gets closed with automation and with fewer people watching more machines. Both of those raise the value of a system that warns you before a line stops.

Bilingual delivery, local-first architecture, IMMEX-aware data handling.

Santa Teresa, New Mexico

Rail-served industrial land, the Union Pacific intermodal complex, and a new 272-acre park with engineering design starting in 2026. Fast-growing sites frequently have modern equipment and no instrumentation layer at all, which is the cheapest possible time to add one.

I worked at CN Wire in Santa Teresa as an electro-mechanic. I know this park.

El Paso, Texas

A 250-acre advanced manufacturing district, a federally funded aerospace and defense cluster with UTEP, Fort Bliss and White Sands, and new plants in drones, autonomous systems and medical devices. Defense and medical work carries traceability requirements that suit an evidence-first approach.

Also where the regional data center question is being decided. See the standing record.

// 005 · Background

The receipts.

  1. 2024 to present

    Celaya Solutions Research · Founder and Principal Researcher

    Independent lab building multi-agent and retrieval systems. Seventy-nine catalogued instruments, fifteen with published source excerpts and measured results. Conservative portfolio valuation of 3.0 to 6.1 million dollars across seventeen proprietary instruments, using COCOMO II methodology.

  2. 2025 to 2026, and 2016 to 2017

    Schneider Electric · Wiring Technician, Quality Assurance · EP44, El Paso

    Quality assurance on 38 kV gas-insulated switchgear with insulated busbar, and 480V low-voltage distribution panels for hyperscale data center and mission-critical applications. Drove a 20 percent improvement in quality outcomes through defect identification, root cause analysis and assembly process correction. First-article and rework inspection prior to factory release.

  3. 2021 to 2022

    T5 Data Centers · Critical Facilities Technician II

    Operated and maintained critical infrastructure at a hyperscale data center. Held 100 percent uptime on critical systems through CMMS-driven preventive maintenance and adherence to Methods of Procedure and Standard Operating Procedures. Contractor and vendor oversight against critical facility standards. Switchgear, uninterruptible power supply, generators and cooling.

  4. 2018 to 2021

    CN Wire Corporation · Electro-Mechanic · Santa Teresa, New Mexico

    Troubleshot electrical systems, HMI panels and sensors in a continuous manufacturing environment. Maintained cooling, compression and emulsion systems critical to production. Root cause analysis on equipment failures with corrective actions to reduce recurrence.

  5. 2013 to 2016

    Delfingen and Dal-Tile · Maintenance Technician

    Preventive and corrective maintenance across extrusion, pneumatic, hydraulic and electrical systems. Industrial presses, electrical panels, spray dryers, ball mills and conveyor systems.

Education and certification

  • The University of Texas at El Paso, Electrical Engineering: Fields and Devices (coursework, 3.5 GPA)
  • El Paso Community College, Computer Science: Cybersecurity (coursework, 3.7 GPA)
  • NFPA 70E, Electrical Safety in the Workplace
  • OSHA 10
  • Eleven-plus years equivalent experience in industrial electro-mechatronic and critical infrastructure work

Selected outcomes

  • MORTEM ranked 186 of more than 700 entries at the Colosseum Agent Hackathon
  • EPPE, the region's first civic blockchain instrument, drew editorial inquiry from the New York Times Texas bureau and El Paso Matters
  • Earned Autonomy Architecture preprint, co-authored with Renato Kuipers, prepared for arXiv
  • Celaya Chain Protocol draft published to ethresear.ch
  • Trinidad: 5,435 documents and 1,065,357 passages indexed and searchable

// 006 · Capability

What I can actually pick up and use.

Listed plainly, because people looking for this work usually search for a specific thing. If what you need is not here, the honest answer is probably that I would have to learn it, and I will tell you that instead of quoting for it.

Critical infrastructure

Medium-voltage 38 kV gas-insulated switchgear. Low-voltage 480V distribution for data center applications. Hyperscale critical facilities operation. CMMS-driven preventive maintenance. Methods of Procedure and Standard Operating Procedures. HMI. SCADA. Uninterruptible power supply systems, generators, cooling. NFPA 70E. OSHA 10.

Quality assurance

Defect resolution. Root cause analysis. Assembly inspection. Point-to-point and wiring schematic interpretation. First-article inspection. Contractor and vendor oversight. Bill of materials compliance review. Non-conformance closeout.

Applied intelligence

Multi-agent systems. Model Context Protocol servers. Retrieval augmented generation and vector search with Qdrant, pgvector and FAISS. Fine-tuning with LoRA. Deep Q-network reinforcement learning. Whisper speech to text. Kokoro text to speech. Semantic embeddings. Server-sent events and gRPC streaming. Federated vector queries. Anomaly detection with IsolationForest. Lead-time estimation.

Models and runtimes

Anthropic Claude API. HuggingFace. Ollama. llama.cpp. PyTorch. TensorFlow. LangChain and LangGraph. MLX on Apple Silicon. CUDA. H100 and A100 class hardware. Local-first and air-gapped deployment.

Languages

Python. JavaScript and TypeScript. Swift. Rust. Go. Solidity. C. Java. SQL. Bash.

Application frameworks

Next.js. React. React Native. Node.js. Express. FastAPI. Streamlit. SwiftUI.

Data and infrastructure

PostgreSQL. MongoDB. Redis. SQLite. IndexedDB. Supabase. Parquet. Write-ahead log replication. Docker. FastMCP. WebSockets. MQTT. Modbus. AWS. Azure. Vercel. Railway. Cloudflare. Solana. Stripe.

Delivery

Bilingual English and Spanish documentation and training. Evidence-cited outputs. Provenance chains. Human review gates. Handover so your team can run it. Honest scoping, including telling you when a project is not worth doing.

// 007 · Questions

The ones I actually get asked.

What does an applied intelligence engineer actually do for a factory?
Three things, usually in this order. First, get the data out: most plants have telemetry trapped in a controller or a historian nobody queries. Second, make it answerable: retrieval over drawings, manuals, work orders and quality records, with citations back to the source page. Third, put a prediction on it: fault detection with enough lead time that a repair moves from unplanned downtime into a scheduled window.
Do you work with maquiladoras in Ciudad Juárez?
Yes. Delivery is bilingual, and the systems are built local-first so operating data can stay on site rather than crossing a border to a third-party interface, which for many IMMEX operations is a compliance question before it is a technical one.
Do I need to send my data to a cloud provider?
No. Local-first is the default here, not an upgrade. Inference, embeddings and retrieval can run on hardware you own, which matters when the data is production yield, defect detail, or anything covered by a customer agreement.
How small a business is too small?
None. A two-truck company losing after-hours calls has a real, measurable problem worth solving, and the work starts with a paid audit rather than a retainer.
What makes this different from hiring an agency?
Twelve years of the industrial side. Most people building these systems have never held a meter on energized gear, read a point-to-point drawing, or been on shift when a critical system failed. That background is why the systems here assume failure, keep a record, and stay out of the control path.
Can you teach my team instead of just building for them?
Yes. Every engagement ends with the team able to run and change what was built, and standalone training is available for maintenance, quality and operations groups who want their own people capable rather than dependent.
What does a first engagement look like?
An audit, paid and fixed price. You get a written finding with what is worth building, what is not, what it costs and what it saves. If nothing is worth building, that is the finding, and you owe nothing further.

Start with the smallest thing that proves it.

Most work here starts with one audit rather than a contract. Tell me what is costing you time or money, and I will tell you whether this is the right tool for it.

Start a build hello@celayasolutions.com See what has been built