Multi-agent assistant for a regulated investment platform. Layered architecture with durable orchestration, supervisory agent graphs, and self-hosted LLM inference.

An AI assistant for retail investors built on a layered multi-agent architecture. Temporal handles durable orchestration and human-in-the-loop checkpoints, LangGraph manages supervisory agent graphs, and PydanticAI provides type-safe agent definitions. Self-hosted vLLM on GKE (H200 GPU) handles sensitive client data, routed through Bifrost AI Gateway to external APIs. Deterministic context layer (MCP servers in Go and Python) separates financial calculations and compliance logic from LLM reasoning. Langfuse provides full observability: traces, evaluations, and cost analytics via OpenTelemetry.

Internal CI/CD operations bot connecting Slack to GitLab pipelines. Build triggers, release management, Play Store publishing, and crash reporting.

A Python/Flask/Slack Bolt service built in 6 weeks and evolved over 3 years into the mobile team's primary CI/CD interface. Exposes build triggers, release branch management, version bumping, and pipeline status directly in Slack. Extended with Google Play API v3 integration (AAB publish, track promotion, staged rollout, production monitoring) and Firebase Crashlytics integration via an HTTPS JSON-RPC MCP client. Runs on Kubernetes as a single-replica Socket Mode Bolt service with OpenTelemetry distributed tracing spanning Slack through GitLab CI to Gradle builds.

Tool-neutral AI agent policy framework consumed by Claude Code, Gemini CLI, and OpenAI Codex across multiple repositories.

A policy specification framework that gives AI coding agents consistent project constraints without per-session configuration. AGENTS.md files define architecture contracts, coding standards, and never-do rules at workspace and project level. A deterministic rule-routing table maps file globs to dedicated rule documents, ensuring agents load the correct domain constraints before modifying build logic, CI pipelines, or feature code. Includes 24 project-specific executable skills with YAML trigger phrases for automatic selection, and a 4-agent task-proof-loop workflow (spec-freezer, builder, verifier, fixer) with JSON schema-validated evidence artifacts.

Local-first analytics dashboard for AI-assisted engineering. Unified cost, burn-rate, cache efficiency, and waste-detection across 9 coding assistants.

A Rust toolchain that reads transcripts from Claude Code, Codex, Cursor, OpenCode, Pi, Copilot, Xcode CodingAssistant, Cowork, and Amp, then surfaces an interactive dashboard with cost estimates, 5-hour billing-block burn-rate projection, cache efficiency, task categorization, waste-detection grade, context-window tracking, and rate-limit tracking — all running entirely on the user's machine. Three surfaces ship together: a `heimdall` CLI with embedded web dashboard, a `heimdall-hook` sub-second real-time ingest writing per-tool cost on every PreToolUse (~50 ms p99), and an MCP server exposing 9 analytics tools over stdio + HTTP. A native macOS menu-bar app (Swift) bundles the CLI with browser-session import. Storage is local SQLite; no telemetry, no network egress.

Kotlin-native CLI tools that replaced the entire Fastlane/Ruby release pipeline for Android CI/CD.

A set of Kotlin CLI tools built with picocli and Apache SSHD/JGit that replaced all Ruby/Fastlane automation in the Android CI pipeline over 8 weeks. ReleaseManager handles release branch creation, version bump orchestration, and back-merge safety checks. SlackUploader manages artifact upload and Slack notifications. A lint-codequality converter produces GitLab Code Quality reports from Android Lint and Detekt output. The migration unlocked Gradle Configuration Cache across all CI jobs and eliminated all Ruby environment issues.

Android library with Java-level ANR monitor and C++ native signal handler for catching events the JVM monitor misses.

A production Android library for detecting Application Not Responding events at two layers. The Java-level monitor watches for main thread stalls, while a C++ native signal handler (built with NDK/CMake) catches SIGABRT-class events that bypass the JVM entirely. Integrates with ApplicationExitInfo for post-mortem diagnostics and uploads structured reports to Firebase Crashlytics. Feature-flag-gated for controlled production rollout. Requires understanding the JNI boundary, NDK build configuration, ProGuard symbol retention for native binaries, and threading constraints when calling JVM from signal context.

Personal portfolio and technical blog. Astro 6 + React 19 islands, Cloudflare Workers + D1, WebAuthn admin panel.

This site. Built with Astro 6 and React 19 islands architecture, deployed on Cloudflare Workers with D1 (SQLite) for the database. Admin panel uses passkey-first authentication via WebAuthn. Content pipeline is single-source: MDX for blog posts, JSON for projects and experience, with build-time generators producing static TypeScript data files and seed SQL. Designed with anti-AI-slop principles: minimal, credible aesthetic where typography carries the design.

Android app for network connectivity optimization with local SOCKS5 proxy, adaptive DPI evasion, and encrypted DNS support.

An Android application that optimizes network connectivity by routing traffic through a local SOCKS5 proxy with adaptive DPI evasion capabilities. Operates in proxy mode or local VPN redirection mode. Implements adaptive evasion strategies with semantic markers and adaptive split placement. Provides encrypted DNS via DoH/DoT/DNSCrypt, integrated diagnostics with automatic probing and strategy auditing, per-network policy memory with intelligent network handover handling. Native Rust modules handle performance-critical paths via JNI bridges. Includes a user guide generator that creates annotated PDFs from live screenshots.

Self-hosted messenger between you and the internet. Pulls articles, videos, and feeds, summarizes with LLMs, delivers signal — not feed. KMP, web, and Telegram clients.

A self-hosted content reader and summarizer with three first-party clients. The backend (Python, FastAPI) scrapes articles and extracts YouTube transcripts via a multi-provider chain (Scrapling, Firecrawl, Playwright), then generates structured summaries via LLM through OpenRouter. The mobile/desktop client is built with Kotlin Multiplatform + Compose Multiplatform, sharing ~80-90% of business logic across Android, iOS, and Desktop with offline-first delta sync. A React/TypeScript web client doubles as a Telegram Mini App for on-demand summarization from the browser. Features include offline reading, semantic search, channel digests, and MCP server integration for AI agents. Ktor for networking, SQLDelight for persistence, Decompose for navigation, Koin for DI.