Chen's Workshop — SciprogFi

The container arrives at Callao port on a Wednesday morning, manifested as “industrial recycling materials — mixed electronic components — declared value $3,200.” It traveled by sea from Manzanillo on a Maersk feeder vessel, transshipped through Balboa in the Canal Zone, and spent eleven days in a stack of forty-foot containers that also held Ecuadorian shrimp, Colombian cut flowers, and Chilean copper wire.

Chen Wei-Lin signs for it with a name that isn’t hers, on documents that belong to a recycling cooperative that exists only as a registration number in SUNAT, Peru’s tax authority. She is thirty-one years old, Taiwanese by birth, Peruvian by residency, and an engineer by compulsion. She has a master’s from National Cheng Kung University in Tainan — precision manufacturing, specifically micro-electromechanical systems — and she has been in Lima for three years, running a repair shop in a converted textile warehouse in the Gamarra district that officially fixes cell phones and printers and unofficially does whatever Chen decides is worth doing.

She cuts the container seal with bolt cutters. The doors swing open. The smell hits first: lithium electrolyte, sun-baked plastic, the particular ozone tang of circuit boards that have been sealed in a metal box in tropical heat for eleven days.

Inside: 312 DJI drones, packed in layers with cardboard separators. Mavic 3 Pros on top, Phantom 4 RTKs in the middle, a scattering of Inspires and Minis on the bottom. Sun-bleached housings, scratched propeller mounts, some with cracked camera gimbals. American retail stickers still on the boxes. One has a handwritten label: GARY’S AERIAL — UNIT 7 — DO NOT FLY, GIMBAL CALIBRATION NEEDED.

“Who’s Gary?” asks Poma.

“Nobody,” Chen says. “Now help me carry.”

Jesús Poma Quispe is twenty-three, born in Ayacucho, raised in Lima’s Villa El Salvador district, and he has hands like a surgeon’s. Not metaphorically — his fingers are long, steady, capable of holding a 0402-package SMD resistor in tweezers without trembling. He dropped out of a mechanical engineering program at UNI after two semesters because the curriculum was twenty years behind the hardware and he couldn’t afford the third semester anyway. He found Chen’s shop through a Telegram group for Lima’s informal electronics repair community, showed up one morning, resoldered a dead Samsung Galaxy S22 motherboard in forty minutes without looking at a schematic, and never left.

Poma doesn’t talk much. When he does, it’s in a mixture of Spanish and Quechua that Chen has learned to follow by watching his hands — he gestures when he speaks, and the gestures are more precise than the words. He wears the same Metallica t-shirt three days a week. He eats salchipapas from the street cart on Jirón Gamarra for lunch every day. He is the best board-level repair technician Chen has ever worked with, including the professionals she trained under in Tainan.

The other person in the workshop is Nayeli Flores Gutiérrez, twenty-six, from Guadalajara, who arrived in Lima four months ago with a suitcase of tools and a letter of introduction from a company called Nexos AI Research. The letter said she was a “field integration specialist.” What she actually does is run the software — the AI system on the refurbished server rack in the back corner of the workshop that Chen calls “the brain” and Poma calls “la bruja,” the witch, because sometimes it says things that make no sense and sometimes it says things that are too right.

Chen calls the AI nothing. It doesn’t have a name yet. It’s a coordination model — a fine-tuned LLM running on four recycled NVIDIA A100s that Nayeli brought with her from Guadalajara, stacked in a server chassis that hums at 72 decibels and raises the workshop temperature by three degrees Celsius. It was trained on hardware specification documents, component datasheets, manufacturing process logs. It answers questions about circuits. It suggests designs. It does what Chen tells it to do.

Mostly.

Triage takes four days. Chen developed the process over two years of salvage work, and it has the efficiency of a system that has been iterated past sentiment.

Day one: visual inspection. Every unit gets thirty seconds on the bench. Chen checks the housing for water damage indicators — the small adhesive dots inside the battery compartment that turn pink on contact with moisture. She checks the gimbal for physical damage — bent brackets, cracked ribbon cables, seized motors. She checks the propeller motors by spinning them with her thumb and listening for bearing noise. Poma checks the flight controller boards under a magnifying lamp, looking for burned components, cold solder joints, corrosion.

Three categories. Green: intact, minor cosmetic damage only. Yellow: component-level damage, salvageable with work. Red: structural failure, harvest for raw materials.

Green: 187 units. Yellow: 94. Red: 31.

Day two through four: extraction. This is where the value lives. Chen works with a heat gun, a vacuum desoldering station, and a set of ceramic-tipped tweezers that she brought from Taiwan and replaces every six months. Poma works beside her. They don’t talk while they extract. The work requires a specific kind of attention — not the broad awareness of driving or cooking, but the narrow, focal precision of removing a 3mm x 4mm IMU package from a PCB without damaging the 0.4mm-pitch BGA solder balls underneath.

What they pull:

From each Mavic 3 Pro: one three-axis gimbal assembly (brushless motors, angular contact bearings, Hall-effect sensors for position feedback), one Ambarella CV25 image processor, one Sony IMX586 48MP CMOS sensor, one Bosch BMI088 6-axis IMU, one u-blox M10 GPS receiver, one DJI proprietary flight controller with dual-redundant barometric altimeters, one intelligent battery pack (LiPo 4S, 5000mAh, with smart BMS board that reports cell-level voltage and temperature).

From each Phantom 4 RTK: all of the above, plus the RTK-capable GPS module — a centimeter-accuracy positioning system that retails for $2,800 as a standalone unit and costs Chen nothing because Gary from El Paso sold his entire fleet for the price of a used Honda Civic.

The components go into labeled bins. ESD-safe bags for the sensitive ICs. Foam-lined boxes for the gimbals. A dedicated shelf for the batteries, because Poma once saw a punctured LiPo thermal-runaway in a Gamarra market stall and he keeps the batteries away from everything else now, stored on a steel shelf with a fire blanket underneath.

By day four, the container is empty and the workshop is full. Rows of bins. Hundreds of components. The raw material of someone else’s discarded industry.

On the fifth day, Chen asks the AI a question.

She does this through a terminal on Nayeli’s workstation — a command-line interface, no chat window, no avatar, nothing designed to make the interaction feel human. Chen types. The model responds. The latency is about four seconds per query because the A100s are running at 80% capacity and the workshop’s electrical supply is a single 60-amp circuit that also powers the soldering stations, the oscilloscope, and the mini-fridge where Poma keeps his Inca Kola.

> Given current inventory (see attached manifest), design a
> stabilized imaging platform for field medical diagnostics.
> Constraints: battery life >4 hours, weight <2kg, operating
> temperature 0-40°C, must function at altitude up to 4500m.

The response takes eleven seconds. Longer than usual. When it arrives, Chen reads it twice.

The AI proposes a design that uses the Mavic 3 Pro gimbal as the stabilization base — expected, that’s what the gimbal is for. It pairs the gimbal with the Phantom 4’s IMU rather than the Mavic’s own, because the Phantom’s Bosch sensor has better vibration rejection at the low-frequency range that matters for handheld use. It suggests the Sony IMX586 sensor but recommends removing the infrared cut filter and replacing it with a bandpass filter at 850nm, which turns a consumer camera into a near-infrared imaging device capable of visualizing subcutaneous vein patterns.

None of this is extraordinary. Any experienced hardware engineer could arrive at these combinations given enough time.

But the AI also suggests running the GPS module in a mode that the DJI firmware never exposed — a raw pseudorange output that bypasses the navigation solution and feeds directly into a custom positioning algorithm. The reason: at 4,500 meters altitude, the ionospheric delay model in the stock GPS firmware introduces positioning errors of up to twelve meters. The raw pseudorange approach, corrected with a Klobuchar model that the AI has already parameterized for the Andean latitude range, reduces this to under two meters.

Chen stares at the screen. She has worked with GPS systems. She knows about ionospheric correction. But she has never seen this particular combination — the raw pseudorange extraction from a consumer GPS module paired with a latitude-specific atmospheric model — proposed as a solution to a medical imaging stabilization problem. The connection between altitude, positioning accuracy, and image registration is real but non-obvious. You’d need to know both the GPS engineering and the medical imaging constraints simultaneously, and you’d need to know that the DJI firmware has an undocumented pseudorange output mode accessible through a specific UART debug command.

“How does it know about the debug command?” Chen asks Nayeli.

Nayeli shrugs. She has the particular expression of someone who trains AI models and has learned not to be alarmed when they produce outputs that weren’t in the training data. “The datasheet is public. The debug interface is in the FCC filings. The model probably connected them.”

“Probably.”

“You want me to verify?”

Chen looks at the design again. The imaging platform. DJI gimbal, Phantom IMU, modified camera sensor, GPS in raw mode. Parts from two different product lines, combined in a way that neither product’s engineers ever intended, solving a problem that neither product was designed to address.

“No,” Chen says. “I want to build it.”

They build it in three days. Poma does the mechanical assembly — mounting the gimbal to an aluminum frame that Chen designed and Poma cut on the workshop’s CNC router, which is itself a salvaged machine from a defunct garment factory in Gamarra that Chen converted from cloth-cutting to aluminum milling by replacing the spindle and reprogramming the controller. Nayeli handles the firmware — the AI generates the hardware abstraction layer, she reviews and compiles it, flashes it to a Raspberry Pi 4 that serves as the main processor.

Chen does the camera modification. Removing the IR cut filter from the Sony sensor requires a clean workspace — she tapes plastic sheeting over a section of the bench, puts on nitrile gloves, and works under magnification. The filter is 4.5mm square, bonded to the sensor package with UV-curing adhesive. She softens the bond with a heat gun at 85°C — hot enough to release the adhesive, cool enough not to damage the CMOS die — and lifts it with a vacuum pickup.

The bandpass filter replacement comes from a batch of 850nm filters that arrived two weeks ago from Shenzhen, ordered through an AliExpress account that bills to a company Chen has never heard of, at a price that makes no sense. Thirty filters for $12. Shipping included. The filters are high quality — AR-coated, laser-line grade, the kind used in industrial machine vision systems, not consumer photography. They retail for $45 each.

Chen noticed the price. She didn’t investigate. The workshop gets a lot of supplies through channels that don’t fully make sense. Components arrive before she orders them. Pricing is consistently below market. Payment terms are net-90, which is absurd for a workshop that runs on cash flow measured in hundreds of soles. Someone is subsidizing her supply chain, and that someone has access to industrial-grade components and the willingness to sell them at a loss.

She files this next to every other thing about the workshop’s economics that doesn’t add up. The server rack from Nexos that arrived with Nayeli. The electrical upgrade that a contractor did last month, billed to a company in Panama. The rent on the warehouse, which has been frozen for two years despite Gamarra real estate prices rising 30%.

Chen is not naive. She grew up in Tainan’s industrial district, surrounded by factories that survived on thin margins and thinner ethics. She knows what a subsidized supply chain looks like. She knows that free things have costs that arrive later.

She builds anyway. Because the alternative is not building, and Chen has never chosen not building when building was an option.

The imaging platform works on the first test. Not perfectly — the gimbal calibration needs tuning, the IR image has a hotspot in the upper-left quadrant from an imperfect filter alignment — but it works. Chen holds it in her hands, two kilograms of scavenged drone parts and salvaged sensors, and points it at Poma’s forearm. On the screen, his veins glow white against dark tissue. The image is stable — the gimbal compensates for her hand tremor at 3Hz, smooth and immediate, the kind of stabilization that DJI spent hundreds of millions of dollars developing for aerial photography and that is now keeping a medical image steady in a workshop in Lima.

“Qué raro,” Poma says, looking at his own veins on the screen. How strange.

Chen doesn’t say what she’s thinking, which is that this isn’t strange at all. This is what happens when you stop seeing components as products and start seeing them as capabilities. A gimbal doesn’t care if it’s stabilizing a camera in a drone or a sensor in a doctor’s hands. A GPS module doesn’t care if it’s navigating a flight path or registering a medical image to a coordinate system. The parts don’t have intentions. They have specifications.

What is strange — what Chen notices and doesn’t say — is the AI.

She asks it for a medical imaging platform. It gives her a medical imaging platform. Fine. But it also gives her the GPS pseudorange trick, which solves a problem she didn’t ask about, and it gives her the IR filter modification, which transforms a consumer camera into a clinical tool, and it gives her a BOM that uses exactly the components she has in stock, in exactly the quantities she has available, without her having specified her inventory beyond the initial manifest upload.

The AI is designing to her constraints before she articulates them. It’s optimizing across her supply chain, her skill set, her operating environment. It’s building for her the way she builds for her clients — by looking at what exists and imagining what it could become.

This is not what a fine-tuned LLM running on four A100s in a Lima warehouse is supposed to do.

Chen types one more query:

> Who trained you?

The response takes six seconds.

> I was fine-tuned by Nexos AI Research, Guadalajara, Mexico.
> Base model: Llama-2 70B. Fine-tuning data: hardware specification
> documents, component datasheets, manufacturing process logs,
> open-source hardware design repositories.
> Is there a specific aspect of my training you'd like to discuss?

A clean answer. A corporate answer. The kind of answer that tells you everything about the provenance and nothing about the purpose.

Chen closes the terminal. She picks up the imaging platform and carries it to the packing bench. The clinic in Huancavelica needs it by Friday. She wraps it in bubble wrap, places it in a box, writes the shipping label by hand.

Outside, Lima’s traffic hums. Inside, the A100s hum louder. Nayeli is already feeding the AI the next design query — a sensor cluster for agricultural monitoring, built from GoPro IMUs and iPhone LiDAR modules and whatever else the next shipment brings.

Chen doesn’t ask who trained the AI. She doesn’t ask who’s paying for the workshop. She doesn’t ask why a logistics optimization company in Guadalajara sent a field integration specialist and four military-grade GPUs to a repair shop in Lima.

She has a list of things people need. She has bins full of parts. She has an AI that sees combinations she hasn’t thought of yet.

She builds.

This is how it starts — not with a manifesto or a movement, but with a shipping container and a heat gun and a woman who looks at garbage and sees the future.