# Stack The stack is product-first: workflow, data, backend logic, inference, memory, human review, and the surface that makes the system usable. Same content that the interactive console at /repl types out command by command. ## Build systems - Design: product shape, interaction models, information architecture, design systems, research synthesis, prototyping, usability reviews, visual systems, brand/product identity, Figma, Framer, Sketch, Adobe CC - Frontend: React, Next.js, Astro, TypeScript, modern CSS, product websites, dashboards, review tools, command palettes - Native frontend: SwiftUI, AppKit, macOS and iOS product surfaces - Backend: Node.js, Python, API routes, Cloudflare Workers, server-side product logic, file pipelines, data jobs, internal tools - Data and storage: MongoDB, Postgres, Redis, SQLite/FTS5, GRDB, SwiftData, CloudKit, local-first persistence, searchable memory layers - Product infrastructure: billing hooks, licensing flows, analytics, GTM surfaces, launch loops, operational tooling - AI product logic: agent orchestration, retrieval, ranking, memory, evidence quality, human review, confidence, dissent, audit trails ## Apple platform - Inference: mlx-swift on Apple Silicon, Qwen3 family (8B dense, 35B MoE) - On-device LLM: FoundationModels on the Neural Engine - NLP: NaturalLanguage (NLTagger, NLEmbedding, NLTokenizer) - Math: Accelerate / vDSP for embeddings and vector ops - Storage: GRDB + SQLite + FTS5, SwiftData + CloudKit - Audio: ScreenCaptureKit and AVAudioEngine - Speech: Whisper via CoreML, five model tiers, hot-swap (NotchLive) - Translation: Apple Translation, on-device, 20 languages (NotchLive) - UI: SwiftUI + AppKit, Liquid Glass on macOS 26 ## Cloud bridges - CLI subprocess from Swift: Claude Code, Codex, Qwen Code - BYOK API: Anthropic, OpenAI, Alibaba, OpenRouter, Groq, Together, Ollama - llama.cpp xcframework wrapper for portable GGUF inference - Cloudflare Workers, API routes, OpenNext, and Next.js for hosted product surfaces - Provider selection, local/cloud fallback, and BYOK paths ## This site - Astro 5 (static) - TypeScript - Bun - Vanilla CSS, custom properties, color-mix() - IBM Plex Sans / JetBrains Mono / serif - Hosted as a static build, no backend - Hand-written JS islands ## Per ship ### AVE **Signal.** Designed and shipped an operational editing system where AI reads project context, plans work, and executes only through accountable native timeline tools. **Solved.** Unstructured footage, vague creative intent, and multi-format delivery pressure became a controlled workflow: analyze, plan, approve, edit, export. **Built with.** Electron desktop app for macOS, with React, TypeScript, Vite, Tailwind, local project folders, timeline state, deterministic editing tools, local MLX sidecars, transcript and visual analysis, assistant/provider routing, render-worker export validation, signed Mac distribution, updater metadata, and licensing. AVE supports local models, installed CLI assistants, and OpenAI-compatible API providers. The assistant can help reason through the edit, but supported timeline changes still route through AVE's own tools. Source: /ave.md · Page: /ave ### TheHood **Signal.** Published a developer-preview agent runtime with CLI and MCP surfaces, role-bound loops, approval gates, and runtime-owned evidence. **Solved.** Ad-hoc agent chats became governed local loops where models suggest, the runtime enforces, and users approve risky actions. **Built with.** TypeScript runtime for local governed agent loops, with Node, CLI commands, stdio MCP tools, JSON project configuration, provider adapters, role-bound orchestration, approval state, evidence records, protected-path policies, and local run artifacts. Source: /the-hood.md · Page: /the-hood ### Manwe **Signal.** Built a macOS and web AI decision system with role-bound agents, 50+ evidence surfaces, claim audits, dissent, and Future Paths. **Solved.** Single-answer chat became a defensible workflow: research, dissent, claim checks, conviction shifts, Future Paths, and auditable decision records. **Built with.** Native macOS and hosted web decision room built around a structured decision ontology: decision categories, advisor roles, evidence sources, claims, assumptions, risks, forecasts, dissent, actions, Future Paths, and Deeper Story synthesis. A run frames the choice, gathers context, casts role-bound advisors, queries and ranks evidence, pressures the answer with dissent, tracks conviction shifts, and writes an auditable decision record. The web app handles Quick and Pro runs, saved records, reusable seed material, Pro Future Paths, and same-language recordings for non-English decisions. The Mac beta adds local-first depth: local models, custom providers, persistent advisor memory, advisor pool matching, identity briefs, participation history, conviction shifts, Event Radar, and future-path interviews. Stack: Swift, SwiftUI, MLX-Swift, local Qwen models, Apple Foundation Models, NaturalLanguage, Accelerate, GRDB, SQLite/FTS5, SwiftData, CloudKit, llama.cpp, Claude Code CLI, Codex CLI, Qwen Code CLI, OpenAI-compatible inference routes, BYOK provider routing, managed cloud inference, Next.js/OpenNext, and Cloudflare Workers. Source: /manwe.md · Page: /manwe ### NotchLive **Signal.** Shipped private on-device captions and translation for any Mac app, with no meeting bot or account. **Solved.** System audio from any Mac app became live captions and translations without a meeting bot, account, or cloud speech pipeline. **Built with.** Swift native macOS app built around a real-time audio pipeline. ScreenCaptureKit captures system audio from any Mac app; AVAudioEngine handles microphone input and live voice capture. Whisper runs locally through CoreML / WhisperKit, with model download, load, delete, batch transcription, streaming transcription, and runtime model switching across five tiers. Apple Translation handles on-device translation across 20 languages. Stack: Swift, SwiftUI, AppKit, ScreenCaptureKit, AVAudioEngine, CoreML, WhisperKit, Apple Translation, NSPanel, LSUIElement, local transcript and session storage. Source: /notchlive.md · Page: /notchlive ### NotchPad **Signal.** Shipped a local-first Mac capture inbox for notes, clipboard, snippets, voice notes, and Apple Notes. **Solved.** Notes, clipboard history, snippets, voice notes, and Apple Notes moved into one private Mac inbox around the notch. **Built with.** Swift native macOS app built as a local-first capture layer around the notch. The product combines quick notes, clipboard history, reusable snippets, Apple Notes import/sync, file import, sensitive-text protection, and NotchLive voice-note handoff into one searchable private inbox. Stack: Swift, SwiftUI, AppKit, NSPanel, Liquid Glass, NSEvent global monitors, clipboard monitoring, Apple Notes integration, local-first storage, searchable note and snippet data, AES-256-GCM, Touch ID. Source: /notchpad.md · Page: /notchpad ### Signal News **Signal.** Built an iOS reader that clusters related coverage into finite briefings instead of another infinite headline feed. **Solved.** Disconnected headlines became story clusters with source context, connected events, predictions, and a clear endpoint. **Built with.** Swift native iOS app built around story-level news clustering, finite briefings, and local memory. Related coverage from 60+ sources is grouped with NLEmbedding plus DBSCAN, then shaped into story clusters, source maps, connected entities, predictions, ripple-effect timelines, and end-of-briefing debriefs. The product logic treats the story as the unit, not the headline. Stack: Swift, SwiftUI, NLEmbedding, DBSCAN, llama.cpp, local SmolLM3, Apple Intelligence on supported Pro iPhones, optional Claude inference through the user's own API key, on-device RAG, Knowledge Memory, CloudKit sync. Source: /signal-news.md · Page: /signal-news ## Links - GitHub: https://github.com/lemberalla - LinkedIn: https://www.linkedin.com/in/oncelcebeci/ - X: https://x.com/oncelcebeci - Email: oncelcebeci@gmail.com - Studio: https://tinythings.app - Site: https://www.oncelcebeci.com