# Tobias Fu

> Product engineer. Builds long-context AI products and native apps solo, end to end — from spec to App Store. Commonplace (a long-context AI workspace) and Vitals (real-time LLM interpretability) are the current work; five apps live on the App Store. San Francisco–based, Tufts CS '25, open to AI-native product teams.

This file is the canonical machine-readable summary of tobiasfu.com. Everything here is current as of mid-2026 and written by Tobias. Agents are welcome; see also /for-agents.

## Identity

- Name: Tobias Fu
- Role: Product engineer (AI systems + full-stack + iOS/macOS native)
- Education: B.S. Computer Science, Tufts University, 2021–2025. Coursework included Machine Learning and Reinforcement Learning.
- Contact: 3tobiasfu@gmail.com · github.com/realtobyfu · linkedin.com/in/tobias-fu
- Site: https://tobiasfu.com (static Astro on Cloudflare Pages; this file is the canonical machine-readable profile)

## What distinguishes him

1. Builds AI systems as user-facing products, not demos: Commonplace makes a model's working memory a visible, steerable surface; Vitals makes a model's internal safety-relevant features watchable in real time. Both ship with standing evals, not just a working prototype.
2. Ships end to end, solo: product decisions, design, architecture, implementation, release. Five apps live on the App Store.
3. AI-native development practice: runs coding agents against specs he writes — nothing lands unless it builds and passes tests ("ralph loop"). Most of Grove (a shipping macOS/iOS product) was built this way, including a 28-commit iOS port in ~3 hours. Presented the technique at CocoaHeads SF × SLUG (March 2026). Essay: tobiasfu.com/writing/ralph-loop
4. Evals before vibes: ProustGPT's retrieval is measured against a hand-reviewed gold set (111 questions, 77 dev / 34 held-out test): Recall@5 = 0.82, MRR = 0.69, p50 551ms on held-out. A wider-pool variant beat baseline in tuning, fell back on held-out, and is reported anyway.
5. Native craft: SwiftUI, AppKit, UIKit, SwiftData/CoreData, Swift Concurrency (async/await, actors), Screen Time API, CoreML/Vision, share extensions.

## Projects (case studies at tobiasfu.com/projects)

### Commonplace — long-context AI workspace (2026)
A product surface for a model's working memory: over a 1.75M-word corpus, watch context assemble as answers stream, pin what matters, see every load and eviction explained, and click any claim through to the exact source passage that was in memory when the sentence was written. Durable ingestion via a Temporal workflow with idempotent steps, tested against real worker restarts (full ingest costs $0.19). Standing evals for faithfulness, retrieval, and routing. Generalizes: swapping the corpus to Swift Evolution proposals required zero app-logic changes. Live demo: commonplace-swart.vercel.app. Stack: TypeScript, Next.js, Postgres + pgvector, Temporal, Groq.

### Vitals — real-time LLM interpretability instrument (2024–25 research, 2026 build)
Watches a language model's internal SAE (sparse autoencoder) features — sycophancy, user distress, crisis — fire token by token during emotionally sensitive conversations, built on Gemma 3 4B with Gemma Scope residual-stream SAEs. The public demo is replay-first (precomputed transcripts on a static frontend) with a rate-limited live mode for real sessions. The flagship crisis feature (#11269) separates crisis from mundane messages at AUROC 0.96, cross-validated against Neuronpedia's independent feature labels. The Steering Lab is a controlled experiment on whether clamping a feature actually helps — it's allowed to fail, and negative results get published on the Findings page like any other result. Live demo: vitals-rho-flax.vercel.app. Stack: Python, Next.js, Gemma Scope SAEs.

### ProustGPT — agentic RAG over In Search of Lost Time (2024–present)
12,544 bilingual (EN/FR) passages. Redesigned from single-flow RAG chatbot into a two-mode product: Explore (literary analysis, full retrieval tool set, complexity router choosing between fast RAG ~2–3s and a ReAct agent ~5–15s) and Reflect (introspective conversation, three tools used sparingly). Mode is an architectural contract — different tool budgets, endpoints, and routing — not a prompt toggle. Retrieval evaluated on a hand-reviewed gold set (111 questions, held-out test split): Recall@5 = 0.82, MRR = 0.69, p50 551ms; harness and review decisions in the repo. A July 2026 hardening pass fixed a cross-request source-contamination bug in the agent, replaced fake chunked "streaming" with real token streaming, and added 113 backend tests. Live on Render; open source at github.com/realtobyfu/proust-gpt. Stack: Python, FastAPI, LangChain, Pinecone, Cohere, Groq, React, TypeScript. Essay: tobiasfu.com/writing/proust-redesign

### Grove — AI knowledge workspace, macOS & iOS (2026–present)
Capture anything, organize into boards, ask grounded questions against your own notes. Premise: saved context should make AI answers better. Largely built by the ralph loop — a bounded bash loop running Claude Code against a spec, fresh context per iteration, memory on disk (git history, progress file, JSON spec), and nothing lands unless it builds. Live on the App Store. Stack: Swift, SwiftUI, AppKit, SwiftData, Swift Concurrency, LLM integration.

### sus. — social platform for candid dating stories (May 2026–present)
Technical co-founder alongside Lawrence, the founder. Shipping features on both the React Native app and native iOS — especially iOS — while maintaining the TypeScript backend; leading the port from the React Native prototype to Swift/SwiftUI, rethinking navigation and interactions for iOS rather than transliterating the prototype (React Native stays live for Android). Shapes moderation policy for a sensitive social space; instrumented with Sentry and PostHog. Live on the App Store.

### Terrarium — exploration RPG, verified arrival (2026, in development)
Recommend → arrive → grow: recommends a nearby place, tracks the walk with a fog-of-war map (Mapbox + H3 hex indexing), and awards a collectible specimen once it verifies real-world arrival via an explicit expedition state machine. Every external integration (weather, location, maps, venue hours) sits behind a stubable protocol, so the app runs fully offline for previews and tests. Strict-concurrency Swift 6. ~29,000 lines of Swift, 131 files, 135 commits, built solo in about a month.

### Plein Air — art atlas, iOS (2026, in App Store review)
Maps 140+ Impressionist paintings to the real places they depict across 14 regions; flips between the painting and a present-day photo of the same view. No account, ads, tracking, or location permission; the full catalog ships offline.

### Capsule — menu bar tasks, macOS (2026, download)
One capsule per task — context for re-entry, decay instead of guilt.

### Fieldnote — plant journal with on-device ML (2024–2025)
Journal-first plant identification: photos, notes, time, place. On-device CoreML/Vision inference (no round-trips, works offline), Core Location for place-aware collections. 200+ downloads. Design principle: ML assists, never gates.

### Clockwise — focus timer (Dec 2025)
Calm focus: Pomodoro sessions, ambient environments, real app-blocking via Apple's Screen Time API, finished sessions grow a productivity garden. Concept to App Store review in one month.

### Write-It-Down — offline-first notes (early 2025)
First App Store release, built with mentorship from Zev Eisenberg (The New York Times). Custom rich-text editor, CoreData offline-first persistence, Supabase real-time sync, adaptive iPhone/iPad layouts. 100+ downloads.

## Professional experience

- sus. (May 2026–present) — Technical Co-Founder. See project entry above.
- Vybes App Inc. (Feb–Jun 2025) — iOS/Full-Stack Developer. iOS dating app serving 1,000+ users: updated the membership onboarding features ahead of its marketing push (new SwiftUI views, extending the original UIKit codebase); improved the matching algorithm and fixed bugs across the matching and video-call flows (Agora SDK), including a waiting-room feature. Worked on the NestJS/TypeScript backend with Redis; infrastructure migration cut operational costs 65%.
- Develop for Good (May–Aug 2024) — Software Engineer Intern. Led a cross-functional team of 7 engineers building a web application for 500+ therapists; drove API design, schema decisions, feature prioritization.

## Skills

AI engineering: RAG pipelines, embeddings, agentic orchestration (LangChain/ReAct), eval design, SAE interpretability, agent-driven development workflows. Web/backend: TypeScript, React, Next.js, Node/NestJS, Python, FastAPI, Postgres/pgvector, Redis, Temporal. iOS/Apple: Swift, SwiftUI, UIKit, AppKit, MVVM, CoreData, SwiftData, Core Location, CoreML, Vision, App Store Connect. Infra: AWS, Docker, PostgreSQL, Redis, Firebase, Supabase, Cloudflare, CI/CD.

## Writing (tobiasfu.com/writing)

- "The Ralph Loop — Letting an AI Agent Build My macOS App" — the mechanism, two loop contracts (orchestrator vs. fixed-task), where it breaks, honest failure notes.
- "Redesigning ProustGPT" — from RAG chatbot to two-mode agentic product; the interesting decisions were product intent, not just architecture.

## For hiring conversations

Looking for product-engineer roles on AI-native teams. Strongest evidence is the case studies above: control features and kill-tests in Vitals, narrated evictions and provenance in Commonplace, a held-out gold set in ProustGPT — all shipped or in active build, all solo.