Independent industrial software engineer, Shenzhen
Integrators and machine builders can win the machine contract and still be short on software people. I cover everything between “the machine arrived” and “your customer’s people run it every day”. The operator interface is the visible part. Below it are three layers, listed in the order they usually cause trouble.
Not ready for a call? Send the device manual or your client’s spec. Within a day you get a straight answer: whether it’s doable, what concerns me, and roughly what it takes. Send it
Where I am right now: I work 10:00–20:00 UTC: afternoons in Europe and the whole US morning. You get a reply within 12 hours.
NowOperator HMI for an industrial robot, for a North American systems integrator (in progress)
Software for machines, and for the people who build them
You ship 20 machines a year and maintain 20 interfaces. Each model has its own screens, its own quirks, its own commissioning days on site, and every support call starts with “which version is it?” I build one operator interface that configures per model, with PLC, Modbus, OPC UA or MQTT underneath, plus the device integration and a remote support path (how access is controlled). Commissioning gets shorter and your team stops maintaining variants.
I also know ERP. I spent two years rebuilding one (BOM, work orders, batches, traceability) before I worked on machine software. The trade calls this layer shop floor to top floor; in ISA-95 terms it’s Level 2 to Level 4. I do that work, but I don’t sell an MES. My ERP experience is an in-house ERP/WMS rebuild and a machine maker’s BOM, work-order and batch systems, not every product on the market; if your customer runs a specific one such as SAP, ask, and I’ll tell you exactly where my experience ends. So when your customer asks whether the machine’s data can go into their system, you don’t need a second vendor to answer: I look at which system it is and tell you plainly what it takes.
A client actually asked me: “Our field engineer adjusted the tool parameters and renamed the points himself, so what exactly did your work buy us?” My answer · If you’re an integrator or an OEM, there’s a page written for you
“It runs, but the operator can’t run it.” One operator interface configured per model, instead of a screen set per machine. Devices that were never meant to be read, read anyway: Modbus, OPC UA, MQTT, serial, vendor SDKs. Alarm and interlock paths that report failures clearly instead of failing silently.
Live now: an operator HMI for an industrial robot, for a North American integrator. And a public repo where every attempt to make a machine do something dangerous is refused, with all checks running in about a second.
“The shop-floor count doesn’t match the ERP count.” Work orders going down, WIP and batches coming back up, traceability, reconciliation. ERP people don’t touch machines, and machine people don’t know what a work order means to the business. Usually nobody owns this layer. I’ve worked on both sides of it.
93% of manufacturers have MES in place, but only 23% report full integration across ERP, PLM, quality and OT systems. That gap is this layer. (Rockwell Automation report, July 2026, 1,560 decision-makers in 17 countries.)
“You’re not on site, so how do I know it works?” I build against a simulated device first, so the software is finished and tested before the real machine is ready. Mock-mode demos show the whole path working before I have any of your credentials. Improvements are locked in with CI checks so they can’t quietly regress, and every milestone comes with written verification and screenshots.
The first half of this has an industry name: virtual commissioning. Industrial companies quoted by Rockwell Automation put the cost of fixing a machine issue during commissioning at up to 100× the cost in design, so I quote this up-front work as a separate line item. Examples: the ClickHouse case below is measured by the database’s own query_log and protected by a CI check, and everything in the demo library runs without credentials.
If your problem isn’t one of these, describe it in your own words. That includes “AI wrote the code for this machine and now nobody dares touch it”, which I hear often and do take on. I’ll tell you honestly whether it’s mine to solve, and if it isn’t, I’ll say so rather than take the job. If the question is whether it can be done at all, there’s a three-day sprint for that
Tools I use: PLC · Modbus · OPC UA · MQTT / Sparkplug B · operator HMI stacks · ClickHouse · MySQL binlog · event-driven services.
One-page PDF if you’re passing me on to someone.
Each project below comes with something you can check yourself: a running demo, measured numbers, or handover docs. The first is the company where I spent nearly two years.
Sole developer for every internal system at Standard Robots, a Shenzhen AMR maker: BOM and supply chain, CRM, deployment, and a binlog recovery that got back 100% of deleted data. Not their fleet software. Signed CTO letter available on request.
100% of wiped production data recovered, by reverse-engineering MySQL binlogs
≈230×less data scanned: 1.9 GiB → 8 MiB, same query, measured by query_log
Same 50M-row table, naive vs tuned ClickHouse schema, measured with the DB’s own query_log: 1.9 GiB → ~8 MiB scanned. One-command reproduce, CI-guarded, with a 3-min walkthrough.
$ python3 break_it.py [1/11] Position outside physical travel PASS [2/11] Position sensor unplugged mid-move PASS [4/11] Person steps into the working envelope PASS [6/11] Undocumented command injected PASS [9/11] Recovery: can it come back after E-stop PASS … All 11 checks held.
9 / 9dangerous attempts refused, plus 2 recovery checks
A machine commanded over MQTT, with a gate layer that classifies every command by the harm it could cause. The test harness runs 11 checks in about a second: nine attempts to make it misbehave, all refused, and two that confirm it can still stop and recover. Those two matter, because a test suite that only checks refusals would pass even if the machine did nothing at all.
Besides these case studies, there’s a shelf of public demos that run without credentials, each with a note on what it does and doesn’t prove. Browse the demo library
Whether I’m good enough, you can check for yourself. The usual worries are different: I can’t reach him during my hours, my finance team can’t pay him, and if he disappears, I’m stuck. Here are my answers to those three, and to the questions that follow from them.
I work nights by choice. It’s when I concentrate best, and it lines up with your working day. I prefer writing because it’s precise, searchable and works across time zones. We get on a call when a decision is stuck.
The CTO of Standard Robots wrote and signed an assessment of my work there. It’s available on request and I can show it any time: tell me who you are and what you’re evaluating, and you’ll have the signed PDF within 24 hours. If you’d rather talk to him directly, I can set up a call. I check with him first, which usually takes less than a week.
I rebuilt a fragile monolithic ERP/WMS as event-driven microservices and worked alongside the business teams from spec to launch. Downtime failures fell by about 90% and throughput went from 50 to 5,000+ QPS. Talking every day with the people who used the system taught me to build what the business needs, which is often not what the ticket says.
50 → 5,000+QPS after rebuilding a panel maker’s ERP/WMS, downtime down ~90%
The job was extracting tables from PDFs: borderless tables, cells that span several lines, several cells squeezed into one. Open-source tools got about 55% right. We combined two engines and matched each cell to the coordinates of its characters by text similarity, which reached 99.5%. Adding a document-conversion path brought multi-format fidelity to 99.99%. That isn’t comparable to today’s LLMs, but the lesson still applies: careful engineering plus domain knowledge beats a raw model. (I’m still on good terms with the founders.)
I led production database migrations across seven microservices and owned the high-concurrency paid flows. This is where I learned to deliver large, non-standard work in phases, one milestone at a time, without dropping anything.
I built and held everything the company ran on except the robot: BOM and supply chain, CRM, expenses, file storage, deployment. Along the way I saw how a configuration becomes a bill of materials, how one model variant multiplies into part numbers, and where after-sales calls really come from. I also recovered 100% of deleted production data by reverse-engineering MySQL binlogs, with operations still running. Backed by a signed letter from the CTO, available on request.
Full-time on my own since 2026. This is my work now, not a side project.
This site is my permanent base. Projects and client relationships run from here, not from any single platform.
What working with me looks like: terms, paperwork, and how to start.
Fixed scope, tiered quotes, milestone payments. The price I quote is the price. If your budget is tight, say so in your first message: I’ll tell you straight away whether it fits, instead of letting you find out three calls later.
For reference, my current engagement: a first project, delivered, then a second, five-figure project with the same client, now in progress.
Together they give your accountant everything needed to reconcile the payment without chasing anyone.
Scope, milestones and deliverables in writing, signed by both sides before any work starts.
One PDF invoice per milestone with your company name, the amount and the project, ready to file.
Pay the way your finance team prefers. If you’d rather use Upwork, we contract there and Upwork’s records apply; if we met on Upwork, we stay on Upwork, since that’s their rule and I follow it. Paying directly works too: through Payoneer, by corporate card or bank transfer, and you receive Payoneer’s official payment record.
US clients: I provide a W-8BEN for your tax reporting. You deal with one person, and the paperwork works like any proper vendor’s.
If your procurement or compliance team needs to contract with a company rather than an individual, tell me and I’ll arrange it before we sign.
First time buying software this way? Most clients book it under project implementation or fold it into commissioning; ongoing work goes in the maintenance line. Happy to write the invoice so it lands where it needs to.
Say this: “I know someone who does the software that ships with machines, from operator screens to machine data landing in the customer’s ERP. He works remotely, and the code goes into your own repo.” Then send them boheastill.com, or forward the one-page PDF as it is. It has a line describing what I do, three things they can verify, and how to reach me.
If the question is “can this be done at all”, you can find out without committing to a project. A three-day Feasibility Sprint gives you a written answer either way, with a running demo if it’s feasible. It costs the same whichever way the answer goes, and if I can’t reach a conclusion you pay nothing. If you already know what you want built, skip it and ask me for a quote; quotes are free.
You paid for a refactor and nothing looks different: same paths, same screens. My answer is not to review the code, but to take a break-it checklist to the machine. If you can’t make it fail unsafely, you’ve seen what you paid for.
You want “an HMI”, and that’s the whole spec. I split the work into milestones so it can’t spiral, and run each one with a single email. From your side, one big risk becomes a series of small stages you can watch being completed.
A quote is cheap and a claim is easy. Giving you something that runs on mock data, before you hand over any credential, moves the risk from you to me.
Pick a slot. You don’t need to write an email first. By the end of the call you’ll have three things: your problem described back to you so you know I understood it, my view on which layer it’s stuck in, and a rough price range. No slides, no discovery phase, no follow-up sequence. If it’s not a fit, I’ll tell you on the call.
The calendar shows times in your timezone automatically, and the invite arrives with a Google Meet link. Slots cover European afternoons and the US morning.
Tell me what your platform can’t do and what that’s costing you. I’ll reply with an honest view on whether it’s solvable, and if it is, usually with a demo.
Don’t want to write an email? Leave your problem and a way to reach you (email is best). You’ll get a reply within 12 hours.