AI agentsIndustry · Facilities · Finance
Mission Copilot
An offline AI app for people who work on site. Walk, scan, talk, and get a signed report with proof for every sentence.
15 of 16printed pages read every line, offline on an iPhone
Watch the demo See the code on GitHub (opens in a new tab)
Proof you can check:Code on GitHubVideo walkthrough
AI agents
From the site to a signed report
- Read on the iPhoneoffline, the AI only picks, it never decides a fact
- Checked by codea value the AI guessed gets caught
- Approved with Face IDsigned in the Secure Enclave, can't change after
- Checked by the serversignature, evidence chain, file hashes
- Endorsed with Touch ID
- Sent back for version 2
Lubo Bali · lubobali.com
The problem
People who check projects on site (roads, buildings, homes, solar, insurance claims) often work far from internet. And the manager back in the office must trust a report they did not see being made.
What I built
An iPhone app that works fully offline, with the AI running on the phone, plus a dashboard for the manager. The officer walks the site, scans paper, talks a short debrief and approves the report with Face ID. Every sentence points to its proof.
Tech & why
The AI runs on the iPhone with MLX: no cloud, no data leaves the phone. Code writes every fact and every sentence, the AI only picks. Face ID signs in the Secure Enclave, so a report can't change after. The server checks every signature.
- iOS (Swift 6)
- On-device AI (MLX)
- FastAPI
- Postgres
- Face ID / passkeys
The result
Tested on an iPhone 13 Pro Max in airplane mode, then the whole chain recorded on production. Printed paper: 15 of 16 pages read every line, 8 of 8 facts right. About 2,800 automated tests.
Watch the demo
What it does
- Works fully offline. The AI runs on the iPhone: no cloud, no data leaves the phone.
- Red flags computed by code, not guessed by AI. For example: 85% paid, 50% built.
- Guided site walk. Photos and location kept as signed evidence.
- Scans paper invoices and reports. Every fact points to its place on the page, and code catches a value the AI guessed.
- Voice debrief. Notes written in your own words, linked to the recording.
- Ask the mission, offline. Every answer cites its source.
- The report is written by code, every sentence with its proof, approved with Face ID (Secure Enclave). It can't change after.
- Manager dashboard: the server checks the signature, the manager opens the proof (photo, scanned page, even the voice), endorses with Touch ID or sends it back for version 2.
- Tamper-evident audit log: "Chain intact".
Built like production
- Test first, about 2,800 automated tests, plus mutation checks.
- CI on every change, 100+ merged pull requests.
- Threat model: 83 threats, each one pointing to a test.
- Staging, then production on a real server.
- Encrypted backups, restore-tested every week.
Tested for real
- On an iPhone 13 Pro Max in airplane mode.
- Printed paper (flat, tilted, low light, creased): 15 of 16 pages read every line, 8 of 8 facts right.
- 39 prompt-injection attacks, 6 of them written knowing my own defenses. Without the checks 8 got through, with them only 2.
- Those 2 are a "perfect forgery" no reading of paper can catch. That's why a person approves every report.
Fits many jobs
Any job done on site, away from the office:
- Field supervision of development projects
- Home inspections
- Construction sites
- Insurance claims
- Property management
- Solar and utility work
If your team works on site and wants something like this, message me.
How this agent is built
Every AI agent has the same five parts. Here is what each one is in Mission Copilot, so you can judge any agent the same way.
- ModelA small AI model running on the iPhone with Apple MLX. No cloud, it works in airplane mode.
- HarnessGuided steps on the phone: download a signed mission pack, site walk, scan, voice debrief, ask, report. Code writes, the AI only picks.
- GuardrailsRed flags by fixed rules, every fact checked against its place on the page, 39 prompt-injection attacks tested, and a person approves every report.
- KnowledgeThe mission pack (signed by the server, encrypted, one key per mission) plus what the officer collects on site.
- ToolsCamera and document scan, location, voice recording, Face ID signing, and sync when the phone is back online.
Safe by design
Data stays on the phone, encrypted, one key per mission. Reports signed with Face ID, every signature checked by the server, and a tamper-evident audit log.
All data is fictional. Code: GitHub (opens in a new tab) (all rights reserved).
Similar projects
- AI agentsMarketMind AIA market analyst agent: live prices, news mood and signals, answered in plain English2.6 s average answer, up to 4 tool calls
- AI agentsSentinel-AIAn AI agent that sorts your inbox, and can never send or delete a thing4 safety layers before the AI touches an email
- AI agentsLubo AI PublisherAn AI content pipeline that writes in my voice from real sources, and waits for my one tap9 posts a week, live since June 2026
Have a problem like this? Get your free AI blueprint in about a minute.
