Houston

A 100% offline, AI-powered Finder alternative for macOS — plus a Mars Base AI Habitat Controller route that demos how Rover's RAG substrate scales from "find a file on my laptop" to "diagnose an ECLSS fault on Sol 423". Cross-platform desktop app built on Tauri 2 + Rust. Demos in airplane mode.

Built for the GDG AI Hack Milano 2026 — MSI "Cut the Cord" track.

Two routes, one binary

URL fragment	Route	Persona	Killer beat
/	Rover Finder	semantic file search + summarize + note + filename + image-describe	bilingual query → top hit at 69% similarity
/#ares	Houston Mars	greenhouse / survival / repair / voice on a 4-persona stack (KV-cache reuse)	typed-or-spoken "Tray 4 EC sensor at 4.2 mS/cm" → grounded NASA-cited diagnose + on-base parts checklist + 5-step repair procedure

Both routes share the same FAISS index, the same MLX-LM / Ollama generator, the same HOUSTON_PREFIX system prefix — Houston Mars is layered on top of Rover Core RAG, not a parallel build. The architecture story is concrete in the response payload (powered_by: rover-core-rag+houston-repair).

┌──────────────────────────────────────────────────────────────────────────┐
│ Houston                                          local · offline · private │
├────────────┬────────────────────────────────────────┬────────────────────┤
│  Folder    │  ⌕  presentazione budget alpha         │  Selected          │
│  ~/demo-…  │                                        │  alpha-budget-Q3.md│
│            │  📝 alpha-budget-Q3.md          69%   │  …/projects/…      │
│  [Index]   │  ~/demo-rover/projects/alpha-bud…     │                    │
│            │  ███████████████████░░░░░░░░░░         │  Reveal · Open ·   │
│  Try:      │  Project Alpha enters Q3 with a        │  Summarize · Note  │
│  budget…   │  revised budget of € 184,500…         │                    │
│  recipe…   │                                        │  - Total budget…   │
│            │  📝 alpha-budget-Q3.md          66%   │  - Capex/Opex…     │
│            │                                        │  - Vendor X risk…  │
├────────────┴────────────────────────────────────────┴────────────────────┤
│ ● sidecar: ready · localhost only      gemma3:4b · nomic-embed · FAISS  │
└──────────────────────────────────────────────────────────────────────────┘

Why local

Pillar (rubric weight)	Houston's pitch
Ottimizzazione tecnica (30%)	Native Tauri 2 + Rust shell (~20 MB binary), single-machine, zero cloud
Utilità pratica (25%)	Find any file by meaning, not filename. Summarize. Generate notes. Open in Finder.
Integrazione creativa (25%)	Native folder picker, reveal-in-Finder, system file mutations gated by native confirm dialogs
Vantaggio competitivo locale (20%)	Airplane mode: zero external requests at runtime. €0 inference cost. Privacy by construction.

Architecture

┌─────────────────────────────────────────────┐
│  Houston.app  (Tauri 2 / Rust shell)          │
│  ┌─────────────────────────────────────┐    │
│  │  React renderer (Vite + Tailwind)   │    │
│  │  SearchBar · FileList · AIPanel     │    │
│  └────────────┬───────────────┬────────┘    │
│               │ invoke()      │ fetch       │
│  ┌────────────▼─────┐  ┌──────▼─────────┐   │
│  │ Rust commands    │  │ Python sidecar │   │
│  │ pick_folder      │  │ (FastAPI)      │   │
│  │ reveal_in_finder │  │ /index /search │   │
│  │ open_file        │  │ /summarize     │   │
│  │ create_note      │  │                │   │
│  │ confirm_move     │  │ FAISS · pypdf  │   │
│  │ move_file        │  │ tiktoken       │   │
│  └──────────────────┘  └────────┬───────┘   │
└─────────────────────────────────┼───────────┘
                                  │ HTTP localhost
                       ┌──────────▼──────────┐
                       │  Ollama (local)     │
                       │  gemma3:4b · gen    │
                       │  nomic-embed-text   │
                       └─────────────────────┘

Process boundaries (deliberate):

• Python sidecar = read-only information layer (parse · chunk · embed · search · summarize). Never touches user folders.
• Rust main = native OS write/dialog layer. Every mutation gated by a native confirm dialog.
• Renderer = stateless UI. Calls Python via fetch, calls Rust via invoke().

Install (end user)

The fastest path: download a signed-by-no-one-but-it-works .dmg from the latest tagged release.

1. Download the latest Houston_*.dmg from the Releases page. (No login required — picks up .app.zip + .dmg from CI.)

2. Open the .dmg, drag Houston into /Applications.

3. First launch: macOS will say "Houston is damaged and can't be opened" or "unidentified developer". This is not corruption — it's Gatekeeper blocking unsigned apps (we didn't pay for an Apple notarization cert). Two ways to fix in 5 seconds:

   Recommended (one-liner in Terminal):

   xattr -cr /Applications/Houston.app
   open /Applications/Houston.app

   This strips the macOS quarantine flag from the bundle and launches it.

   GUI alternative:

   • Right-click Houston.app → Open → Open in the dialog. Or
   • System Settings → Privacy & Security → scroll down to "Houston was blocked from use…" → Open Anyway.

4. Make sure Ollama is running locally with the models pulled:

   brew install ollama
   ollama serve &
   ollama pull gemma3:4b nomic-embed-text

5. Open Rover. First launch creates the Python venv at ~/Library/Application Support/app.houston.demo/backend/.venv (~150 MB, ~1 min). The status footer goes green when ready.

6. Pick a folder via the toolbar's folder picker → click Index folder.

7. Click "Mars Habitat" in the sidebar to see the on-device Mars Habitat AI demo (greenhouse drill-in, voice loop, repair assist).

Voice + STT (optional)

To use push-to-talk to Houston, drop the whisper.cpp model at ~/.local/whisper-models/ggml-base.en.bin (one-time, ~140 MB):

mkdir -p ~/.local/whisper-models
curl -L -o ~/.local/whisper-models/ggml-base.en.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin

Text-to-speech ships natively (/usr/bin/say). No download needed.

Install (developer)

If you want to hack on Rover, clone the repo and run:

./scripts/setup.sh        # pulls models + creates backend/.venv
bash scripts/dev.sh       # launches the live-reload dev stack
                          # (kills zombies, sets LLM_BACKEND=ollama,
                          #  caffeinate so wifi/sleep don't kill it)

Open http://127.0.0.1:1420 (Finder) or http://127.0.0.1:1420/#ares (Mars demo).

Prereqs: macOS 12+, Rust 1.95, Node 22, Python 3.12, Ollama.

setup.sh is idempotent and pulls these models if missing:

• gemma3:4b — generation (~3.3 GB) — default GEN_MODEL
• nomic-embed-text — embeddings, 768 dim (~274 MB)

Build (.app + .dmg) locally

. "$HOME/.cargo/env"
npx tauri build

Outputs:

• src-tauri/target/release/bundle/macos/Houston.app
• src-tauri/target/release/bundle/dmg/Houston_0.1.0_aarch64.dmg

The Python backend/*.py source ships inside the .app as a Tauri resource. On first launch, src-tauri/src/sidecar.rs copies it to ~/Library/Application Support/app.houston.demo/backend/ and creates the venv there — no setup.sh needed for end users.

For tagged releases, build-mac.yml auto-publishes both files to the GitHub Release page. To cut a new release:

git tag v0.1.0-demo && git push origin v0.1.0-demo
gh run watch        # watch the macOS runner build (~10 min)
gh release view v0.1.0-demo

Demo script (3 minutes for judges)

Demo corpus lives at /Users/fp/demo-rover/ (8 files, ~10 chunks):

demo-rover/
├── projects/
│   ├── alpha-budget-Q3.md          ← target of the headline query
│   ├── alpha-notes.md
│   ├── alpha-contract-draft.docx
│   └── beta-roadmap.md             ← semantic distractor (different project)
├── meetings/
│   ├── standup-2026-05-08.md
│   └── retro-2026-04-30.md
└── random/
    ├── ricetta-tiramisu.txt        ← Italian distractor
    └── vacanze-2025.md             ← bilingual distractor with budget tables

1. Open Houston. Footer dot turns green when sidecar is ready (~2 s warm).
2. Pick ~/demo-rover/ via the native folder picker.
3. Click "Index folder" — 8 files / 11 chunks / under 1 s.
4. Type "presentazione budget progetto alpha" — hit Enter.
   • Top hit: alpha-budget-Q3.md at score 0.69. Italian query, English doc — semantic match across languages.
5. Click the result → Summarize. 6 s on warm gemma3:4b. Markdown bullets in the AI Panel.
6. Click "Create Note" — a summary-of-alpha-budget-Q3.md-…md lands next to the source. Confirm via the macOS Finder.
7. Click "Reveal in Finder" — Finder window opens, file selected.
8. Toggle Wi-Fi off and re-do steps 4–7. Identical behavior. Show ./scripts/airplane-check.sh returning 0.

Mars Habitat AI · open /#ares

The same window also runs an on-device Mars Base AI Habitat Controller — the Cut the Cord physics-enforced offline argument: 78 million km from the closest data center, 14-day blackouts every 26 months, Earth-Mars one-way light time 4–24 minutes. We didn't choose offline; Mars chose for us.

Live numbers — MLX-LM Qwen2.5-3B-4bit + tile cache + code-as-action (measured on the demo machine, M3 Pro 18 GB):

Metric	Naive baseline	Optimized stack	Win
Decode tok/s (RAG prefill, 100 tokens)	43.1 tok/s (Ollama gemma3:4b)	57.6 tok/s (MLX 3B)	+34 %
TTFT (RAG prefill, 100 tokens)	2249 ms	1668 ms	−26 %
Greenhouse + procedure A2A (warm)	8124 ms	4050 ms	2.0× speedup
Sensor cache trace 50d→20d→70d	2104 ms	72 ms	29.2× speedup
Code-as-action python_exec	(would need 2nd LLM round)	438 ms	LLM answers quantitative Q without 2nd decode
RAM under load	10.2 / 18 GB	9.09 / 18 GB (50 %)	9.86 GB headroom
Outbound packets during demo	0	0	airplane-check exit 0

Endpoints (additive under /ares):

• POST /ares/houston/greenhouse — narrate per-tray status, cite Veggie / APH chunks
• POST /ares/houston/greenhouse/stream — same, streamed via SSE (TTFT < 2 s)
• POST /ares/houston/survival — habitat / ECLSS tip + severity (NASA-STD-3001)
• POST /ares/houston/repair — the killer query. Free-text fault → diagnose + parts_needed + parts_missing (cross-checked against on-base inventory) + 3-5 step procedure + optional macOS-say TTS
• POST /ares/voice/houston — whisper.cpp + LLM + macOS say round-trip
• POST /ares/voice/houston/text — same pipeline minus ASR (demo-friendly text path)
• GET /ares/sensor/query — tile-lattice cached (29× speedup)
• GET /ares/perf — psutil snapshot + active backend label

Demo flow (90 s, all offline):

1. Open http://127.0.0.1:1420/#ares (or click the route toggle in the header)
2. Click HABITAT → Inventory drill-in (6 bars + 4-crew roster + Houston survival tip)
3. Click 🛠 HOUSTON REPAIR ASSIST (bottom-right amber button) → modal with 3 quick-pick faults
4. Send "Tray 4 EC sensor reading 4.2 mS/cm — fertilizer overdose"
5. Houston replies in ~3.8 s warm: WATCH severity, diagnosis with [S1] citation that opens the actual NASA Veggie PDF in macOS Preview, parts checklist (✓ available / ✗ REORDER NEXT RESUPPLY), 5-step procedure, optional spoken summary via say

Reproduce the benchmark plots (light-theme, no sudo, no internet):

. backend/.venv/bin/activate
python benchmarks/houston/make_light_figures.py
# → benchmarks/houston/out/light/{throughput,cache_lattice,houston_latency,
#                                 voice_breakdown,perf_timeline,a2a_kv_cache}.png

WhatsApp / Slides bundle of all 7 figures: doc/plots-whatsapp/light/ (plus a REPRO_PROMPT.md you can paste into another Claude / Codex on any other machine).

The 2-page tech writeup (doc/TECHNICAL_WRITEUP.{md,pdf}) defends each of the four MSI rubric sub-criteria — model selection, quantization, memory management, inference speed — with a measured number per claim.

Airplane mode proof

./scripts/airplane-check.sh
# → OK — no external network references in source.

Greps every .py / .ts / .tsx / .rs source file for https?:// outside the localhost / Tauri schema allow-list. Exit 0 means clean.

Tech stack & footprint

Layer	What	Footprint
Shell	Tauri 2.11 · Rust 1.95 · WebKit	~20 MB
Renderer	React 18 · Vite 6 · TypeScript 5 · Tailwind 3	~200 KB
Sidecar	Python 3.12 · FastAPI · uvicorn · faiss-cpu	venv ~150 MB
Models	gemma3:4b (Q4_K_M) + nomic-embed-text via Ollama	~3.6 GB on disk · ~3 GB resident

Tests

# Python sidecar
cd backend && . .venv/bin/activate && pytest -v
# → 14 passed

# Rust (smoke)
cd src-tauri && . "$HOME/.cargo/env" && cargo build
# → Finished `dev` profile

# Airplane mode
./scripts/airplane-check.sh
# → OK

Out of scope (deliberate)

File watcher · auto-tagging · batch rename · TTS/STT · multi-folder simultaneous indexing · drag&drop · cloud anything · authentication · multi-user · code signing / notarization · PyInstaller standalone sidecar.

These are all plausible features that would dilute the demo. Houston does one thing — semantic file search with local AI actions — and does it properly.