I build the software that ships with machines β everything between "the machine arrived" and "your customer's people run it every day." Proof before promises.
I take ambiguous, high-risk work from a working prototype to measured delivery and handover. Proof before promises β working demos before contracts.
7 years of Java/Go back-end before industrial Python β I speak ERP as fluently as Modbus.
Public repositories below include open-source tools, independent benchmarks, credential-free demos, and production-derived work. Client work is anonymized where contracts or NDAs require it; each public demo states what was measured, what data is synthetic, and what the result does β and does not β prove.
- Software that ships with a machine β operator HMIs, device integration over Modbus/OPC UA/MQTT, and getting machine data upstream into an ERP. One interface across every model you ship, not a different one per machine.
- Systems that don't talk β the count your line reports and the count your ERP received don't match, and someone spends days each month finding out why. Integrations, protocol bridges and reconciliation.
- Software an AI wrote that won't run β a prototype that demos beautifully and cannot be deployed, extended or debugged by anyone. I take it over and make it survive production.
For hire. boheastill.com β fixed scope, milestone payments, you own the source outright.
| Work | What shipped |
|---|---|
| Inside a robot manufacturer | Sole developer for every internal system at Standard Robots, a Shenzhen AMR maker: BOM and supply chain, CRM, deployment, a 100% binlog data recovery. Not their fleet software β I say so on the page. The CTO signed a letter; it is published in full. |
| High-concurrency backends | Event-driven ERP/WMS, downtime failures reduced by ~90%, throughput taken from 50 to 5,000+ QPS, and a production database migration across seven microservices. |
| Voice AI pipeline | A production donation-call pipeline: phone call β transcription β structured spreadsheet row in about 16 seconds, end to end. |
| Project | Evidence type | Measured result |
|---|---|---|
| ClickHouse DWH tuning | Controlled 50M-row benchmark | A filtered query scans ~230Γ less data; the result is one-command reproducible and CI-guarded. |
| RealSense D405 depth toolkit | Hardware-facing synthetic benchmark | Filter/fusion pipeline with a quantitative accuracy-validation harness, ready for real .bag recordings. |
| Regulatory document parser | Public-source validation | 233-article statute β strictly validated JSON with 0 warnings. |
| District-aware route optimizer | Synthetic routing scenario | 218 β 175 km on identical stops while honoring every mandatory stop. |
| German number ASR | Synthetic speech/noise benchmark | Constrained decoding reaches 52% vs 14% at 5 dB, with the test limits documented. |
| MQTT machine safety gates | Adversarial acceptance harness | Eleven attempts to make a machine misbehave; all refused or safe-stopped, enforced in CI. |
| LLM agent payment gates | Adversarial attack harness | Fifteen attempts to move money that shouldn't move; all fail. Three were added after a reviewer broke an earlier version. |
| Sparkplug B host | Spec-conformance scenarios | Fourteen situations that break naive hosts β stale death certificates, sequence gaps, unknown aliases. No maintained open-source Python host existed. |
Full case studies, walkthroughs and working terms: boheastill.com
- Intranet-Chat-Stream β a self-hosted, DB-less stream for moving text and files between your PC, phone and AI agents. One Go binary, bilingual UI, CI and releases.
- qoder-nix β run Qoder IDE on NixOS with one command.
- pdfSplit β bounded parallel PDF rendering: roughly 1,000 pages from 50 minutes to 2.5 minutes.
- hua-mcp β a fleet of MCP servers with no central registry: one MCP = one directory = one port, discovered by scanning. Documents the three root causes that actually killed servers in production.
π« hi@boheastill.com Β· π boheastill.com Β· δΈζ β boheastill.com/?lang=zh