What I'm building next.
Open-source systems I intend to rebuild from source rather than read about. Each one is chosen for a specific capability it forces me to learn — and each of those capabilities points back at automated process control.
Portable voice assistant
Open-source wake-word and on-device speech stacks
Build a self-contained voice interface that runs wake-word detection and speech recognition locally, so the whole loop works without a network round trip.
Capabilities targeted
Why it transfers
Hands-free operation and local-first reliability are the same constraints that govern instrumentation on a plant floor.
Open bipedal robot replication
NVIDIA / Disney Research BDX-style open droid platform
Reproduce a published bipedal robot from open hardware and control code, and get it balancing under a policy I can read, modify, and explain end to end.
Capabilities targeted
Why it transfers
Closed-loop control, sensor calibration, and drift handling are exactly the mechanics behind DCS level, pressure, and flow control.
FabTwin & FabChem
Own build — hierarchical semiconductor process simulator (FabTwin) and an IPA/water separation flowsheet optimizer (FabChem), cross-validated against public fab datasets
Two evidence-first engineering projects built for process/manufacturing New Grad roles, on a 2027-06-30 deadline.
FabTwin is a plasma-etch SPC and fault-detection simulator whose real differentiator isn't the model — it's validating that model against two real public fab datasets (SECOM, LAM9600) with the correct statistics, including a deliberate reproduction of the common evaluation mistakes (like reporting raw accuracy on a 1:14 class imbalance) as a documented contrast.
FabChem is an IPA/water separation techno-economic optimizer cross-checked against two independent literature NRTL parameter sources, Monte Carlo cost uncertainty, and an independent DWSIM solver run.
SPC — chamber pressure, individuals chart
UCL/LCL = μ ± 3σ
Synthetic — illustrates the chart FabTwin's simulator produces, not a validated result. The right half is drawn open on purpose: cross-validation against the SECOM and LAM9600 public fab datasets hasn't run yet (Gate T1: 2026-09-07 – 2026-10-04). At this project's real 1:14 fault/normal class ratio, raw accuracy is structurally uninformative regardless of the run — precision, recall, and F1 are the actual gate.
Capabilities targeted
Why it transfers
This is the direct target skill set for process and manufacturing engineering roles — SPC, capability analysis, and flowsheet optimization are the daily tools of the internship track this site already documents.
FabTwin: 42/42 unit tests passing (local + CI), ruff/mypy clean, simulator core (config/effects/faults/metrics) implemented and pushed. FabChem: repo scaffolded, environment setup still pending. Gate T1 window: 2026-09-07 – 2026-10-04.
Nothing here is finished. Entries move to shipped only when the build runs and the write-up exists — until then they stay listed as intent.