Independent Software & AI Systems Engineer
On-Device AI · Android · ARM64 Linux · Native Tooling · Software Systems

I build systems close to the hardware: local LLM inference on ARM devices, consent-bound Android automation, native toolchains compiled from AOSP source, and Linux desktops running on phones. Self-taught, and built entirely from an Android phone (Termux / PRoot Debian, no PC) — heavy compilation routes through GitHub Actions, and every hardware-dependent claim is checked against one physical reference device before I call it done.

• On-device AI runtimes — llama.cpp-based local inference on ARM64 Android: GGUF models, streaming, KV-prefix reuse, model-integrity verification, CPU/GPU backend routing
• Consent-driven Android automation — Accessibility and Shizuku privileged execution behind explicit user consent, with a hash-chained audit trail
• Native ARM64 toolchains — Android SDK build-tools and platform-tools built from AOSP source for Linux ARM64/glibc, shipped as SHA-256-verified offline artifacts
• Linux on Android — no-root Debian/Xfce desktops with real GPU routes (Zink/Turnip Vulkan, VirGL) and doctor/repair/benchmark tooling
• Bangla-first applications — business ledger and document tools where Bengali is the primary language, not a translation layer

Practical technology and digital support for individuals, offices, and businesses — with the full service catalog available on the Services page.

Project	What it is	Engineering evidence
LAI	Bangla-first local AI + consent-driven Android automation runtime	Real arm64 llama.cpp CPU inference (GGUF, streaming, KV-prefix reuse) with device-measured throughput and TTFT; Shizuku/Accessibility consent boundaries with a hash-chained audit trail; symbolized root-cause diagnosis of an Adreno Vulkan driver crash (vkCmdBindPipeline SIGSEGV) that shaped a fail-closed CPU-default architecture · v0.9.7
GGEN	Android-first creative & document studio (Flutter/Dart)	Pure-Dart core with 143 unit tests and a Flutter shell with 353 widget/controller tests; deterministic text-layout engine with a proven conservation invariant; transactional file persistence with SHA-256 receipts
ADT	Native ARM64 Android development toolchain	Builds Android SDK build-tools/platform-tools from AOSP source for Linux ARM64/glibc (not only Android/Bionic); ships SHA-256-verified offline release artifacts; validates the full pipeline on real hardware — native source → APK → sign → install → JNI load → run · v37.0.0
Ternux	No-root Debian/Xfce Linux desktop on an Android phone	Zink/Turnip Vulkan and VirGL GPU routes; modular Bash installer with doctor/repair/benchmark tooling; glmark2 score 140 (OpenGL 4.6 via Zink) measured on device, with device evidence that separates measured results from untested claims
Songjog	Bangla-first business & institution operations app	Owner Edition: fast daily entry, local SQLite records, auditable corrections instead of destructive deletes; 94 tests green on CI; export/diagnostics validated on the physical reference device

Each repository documents what is implemented and device-verified separately from what is experimental or planned. The current boundary:

• Validated on hardware — arm64 llama.cpp CPU inference with measured throughput/TTFT; AOSP toolchain builds for ARM64 glibc with verified artifacts; Vulkan desktop rendering on Android via Zink/Turnip with a measured glmark2 score; Bangla text shaping and layout invariants under test
• Diagnosed, deliberately not shipped — the Adreno Vulkan crash in LAI's GPU inference path: root-caused with symbols, then gated fail-closed to CPU rather than shipped half-working
• Experimental / in qualification — GPU LLM acceleration (Vulkan on Adreno) and Qualcomm Hexagon/QNN NPU paths: treated as qualification gates, not shipped capabilities, until device evidence exists

• Evidence before claims — separate implemented, measured, experimental, and planned work.
• Fail closed — unsupported or unstable hardware paths stay gated instead of being presented as working.
• Phone-first is a constraint, not a slogan — the development loop is designed around one Android device and remote heavy builds.
• Reproducibility matters — source, build inputs, checksums, tests, and release artifacts should be traceable.
• Consent and auditability — privileged Android automation is explicit, bounded, and logged.
• Offline resilience — important workflows and artifacts should remain useful without a permanent network dependency.

The projects are separate systems with complementary roles: LAI focuses on local AI and Android automation; ADT provides native ARM64 development tooling; Ternux provides the Linux-on-Android environment; GGEN and Songjog turn the platform into user-facing applications; DocDr explores an offline document workspace. The common thread is the same build → deploy → measure → document loop.

Only the heavy build runs remotely, on GitHub Actions. Everything else — writing and editing code, pulling the built artifact back, installing it over ADB, launching it, debugging what goes wrong — happens on the same phone against the same physical reference device. Failures loop straight back into a fix → rebuild → redeploy → retest pass. No separate build machine, test lab, or handoff between roles.

Qualifying GPU/NPU acceleration paths (Vulkan on Adreno, Hexagon/QNN) against real hardware, hardening CI/release pipelines across these projects, and building DocDr — a mobile-first offline document workspace, early development.

More detail — bilingual (বাংলা/English), with an evidence-graded claims model and a per-claim verification log: soobujmiah.github.io