Eltavine
Eltavine

00Software engineer

Android,Windows,native.

I build Android security tools, network infrastructure, and local-first workspaces with clear signals and practical reliability in mind.

A compact portfolio of security tools, platform work, and the technologies behind them.

View projectsGet in touch
Signals
Local checks
Native probes
Reliable systems
A hand-drawn crayon portrait of Eltavine holding a red heart
Fig. 01 — Eltavine, crayon on paper
scroll

01Projects

Featured work.

Selected work across Android security, network infrastructure, and local-first workspaces.

Projects
06
Public
04
Internal
02

Android / Security

Duck Detector

Duck Detector is a device-side Android security inspection app that keeps analysis local.

It turns device trust signals into focused local checks, from bootloader state to root traces and attestation results.

Built with

  • Android
  • Kotlin
  • C++
  • ASM
Highlight in stack
Dwg no.
ED-01
Rev
C
Scale
1 : 1
Sheet
01 / 06
Status
Native Android security tooling
Source
eltavine/Duck-Detector-Refactoring1.1K stars174 forksas of Oct 2026
Date
Oct 2026
Drawn
Eltavine

Public

star history ↓
5001k1.5kAprJunAugOctstars over time
Star history · eltavine/Duck-Detector-Refactoring · as of Oct 2026
ON DEVICEbootloaderroothookmountvirtualizationattestationKotlinchecks · ComposeNativeC++ · ASM probesLocalreportanalysis stays on the device
Fig. 01 — local inspection flow
3 more notes ↓
  1. Checks risky device states locally, including bootloader, root, hook, mount, virtualization, and attestation signals.
  2. Uses Kotlin and Jetpack Compose for a clear Android inspection flow.
  3. Adds native probes where lower-level device evidence matters.

Rust / Networks / Reliability

Pingora Panel

Pingora Panel is a team operations gateway control platform built around a Pingora data plane, versioned configuration, and durable release workflows.

It turns Pingora into a durable control plane for team operated gateways, keeping configuration, activation, rollback, and recovery explicit.

Built with

  • Rust
  • Pingora
  • Tokio
  • Axum
  • Tonic
  • Protobuf
  • Docker
  • Linux
Highlight in stack
Dwg no.
ED-02
Rev
B
Scale
1 : 1
Sheet
02 / 06
Status
In progress / durable gateway foundation
Source
eltavine/pingora-panel1 stars0 forksas of Oct 2026
Date
Sep 2026
Drawn
Eltavine

Public

illustrative sketch ↓
snapshot v13health check failedsnapshot v12active · last known goodsnapshot v11SUPERSEDEDsnapshot v10SUPERSEDEDrolled back
Fig. 02b — release log (illustrative)
Versioned DSLCompilerEngine-neutral IRPorts & adaptersActivationCAS · atomic snapshotsPingoradata planeLast known goodrollback snapshotdrains gracefully,fails closed
Fig. 02 — config pipeline and release safety
3 more notes ↓
  1. Uses a versioned DSL, compiler pipeline, engine neutral IR, and ports and adapters to keep configuration semantics independent from the data plane.
  2. Exercises durable systems behavior through CAS activation, atomic snapshots, Last Known Good recovery, graceful drain, and fail closed management boundaries.
  3. Connects a Pingora data plane to Proto first contracts, Tonic gRPC health, REST adapter foundations, observability, and black box verification.

Flutter / Rust / Local-first

Yurimashi

Yurimashi is a local-first, cross-platform AI workspace with a Flutter interface and a secure Rust runtime.

It brings a secure native runtime, provider adapters, encrypted local storage, and a cross platform Flutter shell together behind explicit contracts.

Built with

  • Dart
  • Rust
  • Protobuf
  • Riverpod
  • SQLCipher
  • Linux
  • Android
  • iOS
  • macOS
Highlight in stack
Dwg no.
ED-03
Rev
A
Scale
1 : 1
Sheet
03 / 06
Status
Internal development
Source
Private repository
Date
—
Drawn
Eltavine

Internal

illustrative sketch ↓
WORKSPACEAGENTSAITTS
Fig. 03b — workspace UI sketch (illustrative) · details redacted
Flutter UIandroid · ios · macos · windows · linuxPROTOBUF CONTRACTS · BUFRUST RUNTIMEAgentcore servicesProvidersAI · TTSStoreCAS · SQLCipherlocal-first, encrypted at rest
Fig. 03 — runtime layers
3 more notes ↓
  1. Organizes an Android, iOS, macOS, Windows, and Linux workspace around Flutter UI, Rust runtime and persistence, and Buf managed Protobuf boundaries.
  2. Treats local first storage, encrypted CAS, SQLCipher, backups, privacy controls, and fail closed capability handling as product behavior.
  3. Provides a modular agent foundation with multiple AI and TTS provider adapters, durable runtime services, IPC, and independently testable package boundaries.

Tauri / Android / Device operations

yunheshiguang

yunheshiguang is a modular Tauri workbench for managing and automating Android device fleets.

It is shaped as a high-density Android operations desk: connect devices, see their live state, and send carefully bounded actions across a batch.

Built with

  • Tauri
  • React
  • TypeScript
  • Vite
  • Tailwind CSS
  • Rust
  • Protobuf
  • ADB
  • scrcpy
  • Vitest
Highlight in stack
Dwg no.
ED-04
Rev
A
Scale
1 : 1
Sheet
04 / 06
Status
Internal development
Source
Private repository
Date
—
Drawn
Eltavine

Internal

illustrative sketch ↓
OPERATIONS DESKSelectShotInstall6 SELECTEDTASK LOG
Fig. 04b — operations desk sketch (illustrative) · details redacted
Tauri shellReact · TypeScript UIGENERATED PROTOBUFRust commandsdevices · groups · tasksADB · scrcpyFLEETone action,the whole batch
Fig. 04 — fleet control path
3 more notes ↓
  1. Builds a Tauri desktop shell with React, TypeScript, Vite, Tailwind CSS, and Shadcn inspired UI conventions.
  2. Models devices, groups, tasks, diagnostics, and localization as separate boundaries, with Rust commands exposed through generated Protobuf contracts.
  3. Covers ADB and scrcpy workflows including device discovery, batch screenshots, recordings, file transfer, app actions, text input, and recovery diagnostics.

C++ / Qt / Desktop messaging

SereinGram

SereinGram is an open source Telegram client for macOS, Windows, and Linux, built on Telegram Desktop and inspired by Nagram and AyuGram.

It adds ghost mode, a local message history, filters, and translation services to Telegram Desktop, while every addition starts switched off and reaches upstream code only through a narrow hook facade.

Built with

  • C++
  • Qt
  • CMake
  • Protobuf
  • Buf
  • Python
  • SQLite
  • Docker
  • Linux
  • macOS
Highlight in stack
Dwg no.
ED-05
Rev
A
Scale
1 : 1
Sheet
05 / 06
Status
In progress / pre-release
Source
eltavine/SereinGram4 stars0 forksas of Oct 2026
Date
Oct 2026
Drawn
Eltavine

Public

illustrative sketch ↓
CHATCHATSGhostDELETEDEDITED
Fig. 05b — chat window sketch (illustrative)
Telegram Desktopupstream call siteONE LINEHooks facadeserein/hooksNO HANDLERUpstream defaultsame as TelegramMODULESGhostHistoryFiltersevery modulestarts switched off
Fig. 05 — one-line hooks into upstream
3 more notes ↓
  1. Reaches upstream only through one line calls into a serein/hooks facade that defaults to Telegram's own behavior, with a CI budget capping how many upstream files and lines change.
  2. Declares settings, structured configuration, export bundles, and history records in proto3 with Buf lint and breaking checks, and generates typed C++ options, codecs, and settings pages from them.
  3. Keeps deleted and edited messages in an AES-GCM encrypted Qt SQL history, and builds macOS universal, Windows, and Linux AppImage, deb, rpm, Flatpak, Arch, and Nix packages in CI.

Kotlin Multiplatform / Reading

Zhihu-Hyperion

Zhihu-Hyperion is an ad-free, lightweight, AI-assisted Zhihu client for Android, iOS, Windows, Linux, and macOS, built on Zhihu++.

It hands the feed back to the reader. Ads, promoted posts, and paywalled columns are filtered out on the device, recommendations can run entirely locally with weights you choose, and nothing is sent as telemetry.

Built with

  • Kotlin
  • Compose
  • Ktor
  • SQLite
  • ONNX Runtime
  • Android
  • iOS
  • macOS
  • Linux
Highlight in stack
Dwg no.
ED-06
Rev
A
Scale
1 : 1
Sheet
06 / 06
Status
Released / stable and nightly builds
Source
eltavine/Zhihu-Hyperion11 stars0 forksas of Oct 2026
Date
Oct 2026
Drawn
Eltavine

Public

illustrative sketch ↓
HOMEFeedFollowHotDailyAIFILTERED · PROMOTED
Fig. 06b — home feed sketch (illustrative)
ZHIHU FEEDansweradarticlepromotedpaywalledvideoFiltersrules · quality · ONNXFEEDfiltered onthe device,no telemetry
Fig. 06 — feed filtered on the device
3 more notes ↓
  1. Splits one Kotlin Multiplatform codebase into core, feature, and shared modules wired with Koin, so the Android, iOS, desktop, and macOS apps share one Material 3 Expressive interface in Compose Multiplatform.
  2. Filters content with regex keywords, blocked users and topics, and quality thresholds, plus semantic blocking from an on-device ONNX Runtime embedding model in the Full build, while the Lite APK stays under 5 MB.
  3. Ships signed Android builds, unsigned iOS IPAs, a Windows MSI, a Linux AppImage, and a native macOS app from CI, each release carrying an update.json manifest, SHA-256 sums, and build provenance.

02About

Notes on how I work.

The short version of who I am, plus the rules I keep coming back to.

hello!

I’m Eltavine, a software engineer who likes to work close to the platform: Android internals, native code, and the infrastructure that keeps services running.

A lot of my work starts with a trust question. Is this device in the state it claims? Is this config safe to roll out? Does this data need to leave the machine at all? I like answering those with tools that check locally and show their evidence.

I care about calm software: clear signals, honest status, and failure modes that are designed rather than discovered.

Working rules

  1. 01Keep it local.

    Checks run on the device, and data stays on the machine by default.

    see it in Fig. 01 ↘

  2. 02Make boundaries explicit.

    Contracts, an engine-neutral IR, ports and adapters: the seams are written down so each side can change safely.

    follow the pipeline in Fig. 02 ↘

  3. 03Fail closed, recover cleanly.

    When something breaks, stop safely, fall back to a last known good state, and treat recovery as a normal path.

    watch the rollback in Fig. 02 ↘

On the bench

03Stack

Tools I reach for.

A short map of the languages, platforms, and services I use most. The colored dots point back to the projects that use them.

Cross-reference

Core

Core Languages

Languages for native code, services, and performance-sensitive work.

  • C++, used in Duck Detector and SereinGram
  • C
  • Rust, used in Pingora Panel, Yurimashi and yunheshiguang
  • Go

Application

Application Languages

Tools for Android apps, desktop clients, web interfaces, and shipped product features.

  • Kotlin, used in Duck Detector and Zhihu-Hyperion
  • Compose, used in Zhihu-Hyperion
  • TypeScript, used in yunheshiguang and This site
  • JavaScript, used in This site
  • Java
  • Qt, used in SereinGram
  • React, used in yunheshiguang
  • Astro, used in This site
  • Vite, used in yunheshiguang and This site

Platform

Systems and Platform

Systems I use for device, server, and deployment work.

  • Linux, used in Pingora Panel, Yurimashi, SereinGram and Zhihu-Hyperion
  • Windows Kernel
  • Android, used in Duck Detector, Yurimashi and Zhihu-Hyperion
  • iOS, used in Yurimashi and Zhihu-Hyperion
  • macOS, used in Yurimashi, SereinGram and Zhihu-Hyperion
  • Docker, used in Pingora Panel and SereinGram
  • Cloudflare, used in This site
  • Ory

Data

Data Layer

Storage, caching, and API tools for product data.

  • PostgreSQL
  • Redis
  • Protobuf, used in Pingora Panel, Yurimashi, yunheshiguang and SereinGram
  • ConnectRPC

Workflow

Build and Governance

Build, package, and schema tools that keep builds repeatable.

  • pnpm, used in This site
  • Turborepo
  • CMake, used in SereinGram
  • Buf, used in SereinGram

04Contact

Direct, verifiable, quiet.

Email stays protected until intentionally revealed. PGP material is kept separate and minimal.

Say hi —

Email
GitHub
github.com/eltavine

— Eltavine

fingerprint seal · visual hash
PGP key
09E517BD5F0084C8
Fingerprint
46CC2913C0B6460FF498DE6009E517BD5F0084C8
Public key
Download .asc