Build research software and verification tools
Experiment and measurement tools, sensor data visualization, performance comparison, cross-OS verification — designed and built around your specific research goal.
Research → Working System
Turning ideas
into working systems.
Working across AI, embedded systems, devices, and the web, I turn research ideas and technical concepts into testable software and working prototypes. I can also join mid-project — investigating why something isn't working, or picking up a technical evaluation already in progress.
Feel free to reach out even mid-way through research, prototyping, or debugging.
40+ YEARSFollowing technology as it changes
CROSS-DOMAINAI / Embedded / Devices / Web
END-TO-ENDDesign, implementation, hardware verification, operation
29 TITLESTechnical book translation, supervision, and editing
What I Can Take On
What I Can Help With
Organized by the kind of problem you can bring me, not a list of technologies. This includes tools that fall outside your core research but still need building, and bugs whose cause isn't known yet.
Connect AI to devices, the OS, and sensors
Using MCP servers and similar bridges, let an AI client operate BLE devices, capture the screen, monitor serial logs, or reach into a business system.
Prototype AI / LLM-powered tools
From multi-agent systems and structured output to full prototypes with auth, a database, and streaming. On-device inference that reduces cloud dependence is also in scope.
Build sensor / microcontroller / IoT hardware
Multi-MCU coordination, wireless communication, low-power design, cloud integration, and companion phone apps. Delivered as working hardware plus handoff documentation.
Isolate why something doesn't work, or runs slowly
I isolate root causes by measuring and testing hypotheses — reading library source, checking OS settings, tracing communication, right down to the hardware boundary.
Selected Work
Selected Work
Each entry states clearly whether it's a concept, a prototype, hardware-verified, or in production. For client work that can't be disclosed, the company name and confidential details are withheld.
Building MCP servers, one per use case, to connect AI to the physical world
MCP server suite
Rust / TypeScript / MCP
I/O adapters that let an AI client reach BLE devices, capture the screen, or monitor serial logs across multiple devices.
Implementation scope
Implemented BLE read/write/notify, screen and window capture, multi-device serial log monitoring and summarization, a web UI, and both stdio and HTTP transports.
● In daily use
TTS x lip-sync framework for AI agents
Tauri / Rust / Azure Speech / MCP
Gives AI dev-support tools like Claude Code a voice and a 3D avatar body, via MCP. Used daily to have TTS read out the progress of dev work.
Implementation scope
An MCP server built into the Tauri app (list voices, speak, play VRMA animations, and more), TTS via Azure Speech Service, VRM avatar lip-sync, and Windows/macOS/Linux support.
End-to-end communication verified on real hardware
Bluetooth PAN / low-power IoT communication
C / ESP-IDF / BTstack
Implemented BNEP/PAN, which isn't in the standard stack, to get internet access via a smartphone. Also designed the wake-cycle correction and current-measurement methodology.
Implementation scope
Verified SDP, SSP pairing, BNEP, DHCP, and HTTPS communication on real hardware. Corrected deep-sleep wake-cycle jitter using EWMA prediction.
Server and client implemented (integration testing pending)
PC control from a VR headset
Rust / NVENC / OpenXR
A remote-desktop system that streams a Windows app's screen over HEVC and displays it as a manipulable panel in VR space.
Implementation scope
Implemented zero-copy GPU encoding, UDP video delivery, input injection, OpenXR rendering, 6DoF manipulation, and hand tracking.
● In production
AI chatbot for LINE
Node.js / nginx / systemd
Runs continuously on a self-hosted server. Supports switching between multiple LLMs, web search and image generation, and an admin panel — hardened against stalls and rendering glitches hit in real operation.
Implementation scope
Implemented the webhook, an authenticated admin panel, switching between OpenAI/xAI/Google, selective replies in group chats, and Flex Message rendering with a plain-text fallback.
Delivered (client project, company name withheld)
Production-line model for demoing a manufacturing management system
Arduino / Raspberry Pi / PLC
A tabletop production-line model built to demonstrate a manufacturing management system — reproducing conveyance, processing, assembly, inspection, pass/fail judgment, and sorting. Integrates role-split coordination across multiple MCUs with image- and color-based judgment.
Implementation scope
Built out motor PWM control, PLC relay communication, photoelectric sensors, camera and color-sensor integration, servo control, a shared library, and handoff documentation.
How I Debug
How I Debug
I consider the process of getting to a root cause — not just the deliverable — part of what makes technical support good.
Don't fix on a hunch —observe first, then decide.
It wasn't the GPU that was slowIsolated and measured stuttering in object detection, and traced the cause to the camera's auto-exposure.
Symptom
Real-time object detection video was stuttering.
Approach
Instead of assuming "the model inference must be heavy," measured GPU inference and camera frame delivery independently.
Measurement
Tested each hypothesis separately: USB bandwidth, sensor readout, warm-up time, driver settings, and a dirty lens.
Conclusion
The GPU was using only 1/22 of its inference capacity — the fix needed wasn't model optimization, it was better lighting.
The servo problem was an OS settingTraced the issue past the wiring and code, into PWM clock management, and resolved it there.
Symptom
A servo connected to a Raspberry Pi wasn't moving to the commanded angle.
Approach
Instead of stopping at wiring and code, investigated the OS's PWM clock management as well.
Measurement
Isolated the default settings and found a hardware clock conflict with the audio driver.
Conclusion
Resolved with a configuration change, and documented the cause and fix in the README to prevent recurrence.
Same panic, a different cause each timeRefused to treat the ESP32 crashes as one bug, and separated them into several independent ones.
Symptom
Repeated crashes during HTTPS communication, always showing the same panic message.
Approach
Before just increasing the stack size, switched to logging remaining stack space and capturing backtraces.
Measurement
Worked through stack monitoring, disabling unneeded features, freeing heap, and backtrace analysis in sequence.
Conclusion
Identified and fixed stack pressure, an object-lifetime issue, and an invalid free during disconnect handling — each a separate bug.
Publications / Editorial Work
Publications
Editorial / Translation / Technical Books
29titles
Technical book translation, supervision, and editing (under the company's name)
Technical book translation, supervision, and editing (under the company's name)
From the late 1990s to the early 2000s, I worked on translation supervision, technical supervision, editing, and writing for PC and programming books. Translating and editing complex technology from abroad into something Japanese engineers could actually use still shapes how I design and build today.
In addition, under my own name, I supervised or translated titles including "Windows 95 Bible."
Windows 98 Bible
Windows 2000 Professional Bible
Active Directory (in depth)
Game Creators' Bible
Professional Java (Vol. 1 & 2)
Introduction to Java Studio
Show all 29 titles under the company name
- Windows 95 Game Programming: Intro to DirectX (Supervising Translator, 1997)
- Game Creators' Bible (Supervising Translator, 2001)
- The Complete Guide to Powerful CGI Techniques (Supervising Translator, 1997)
- Active Directory (in depth) (Supervising Translator, 2001)
- The Complete Guide to Powerful Java Techniques (Supervising Translator, 1997)
- Build Your Own PC: Windows 98 Edition (Editor, 1999)
- Build Your Own PC: Windows Me/2000 Edition (Editor, 2000)
- The Complete Guide to Powerful Perl 5 Techniques (Supervising Translator, 1998)
- The Complete Guide to Powerful Visual Basic 4 Techniques, Vol. 2 (Supervising Translator, 1996)
- High-Speed 3D Graphics for Windows 95/98/NT (Translator, 1999)
- Professional Java, Vol. 1 (Supervising Translator, 2001)
- Professional Java, Vol. 2 (Supervising Translator, 2001)
- C# Language Reference Black Book (Supervising Editor, 2002)
- Practical C# Programming (Supervising Editor, 2001)
- The Complete Guide to Powerful C++ Techniques (Supervising Translator, 1999)
- C++ MIDI Programming (Supervising Translator, 1999)
- Introduction to DirectX 3D Game Programming (Supervising Translator, 2000)
- Introduction to DirectX 5 Game Programming (Supervising Translator, 1998)
- Introduction to Java Studio (Author, 1998)
- The Complete Book of Linux (Supervising Translator, 2000)
- Web Page Sound Techniques (Supervising Translator, 1998)
- Windows 2000 Professional Bible (Supervising Translator, 2000)
- Windows 2000 Professional Networking Bible (Supervising Translator, 2000)
- Windows 98 Secrets (Supervising Editor, 1998)
- Windows 98 Tuning Black Book (Supervising Editor, 1999)
- Windows 98 Bible (Supervising Translator, 1998)
- Windows 98 Bible, Professional Edition (Supervising Translator, 1998)
- Windows 98 Registry Black Book (Supervising Editor, 1999)
- Windows NT Server 4 Network Bible (1997)
About
Background & Current Interests
My earliest technical work was porting Fantavision — an Apple II animation program — to the PC-9801, building the assembly-language animation-generation engine. From there: translation-supervising and technical-supervising books, embedded systems, industrial demonstrator hardware, mobile apps, and AI agents. Rather than stay in one domain, I've kept doing the same underlying work — understanding complex technology, verifying it, and turning it into something other people can use. Right now I'm especially interested in connecting AI to the OS, devices, and sensors, instead of leaving it confined to a chat window.
Connecting AI agents to the physical world
Mechanisms — MCP and similar — that let AI operate devices, OS functions, and business systems.
Cloud-independent inference
Running LLMs entirely on-device, designed to reduce dependence on connectivity, accounts, and external APIs.
Low-power communication grounded in real measurement
Measuring how power-saving modes actually behave, and designing from those measurements rather than datasheet numbers.
Timeline
Development Activity Timeline
From October 2024 to the present, cross-referencing records left in both macOS and Windows working environments (git commit history, READMEs, work logs) with activity on GitHub — personal repos and the cami org (my own team's repos) — laid out chronologically. It shows a throughline from book supervision/translation to AI and embedded work, growing markedly denser over the last two years.
Contact
From the research-and-concept stage
onward — feel free to reach out.
It's fine if you're still at the "I'm not sure what to even ask for" stage, or want to talk through how to verify something before starting.
EMAILsatbah@gmail.com
GITHUBgithub.com/satbah ↗