Inspiration
When a voice chat or game suddenly drops, people often jump from one frightening guess to another: Did someone get my IP? Should I restart everything? Am I still under attack? Search results usually add noise. A generic chatbot can give advice, but it does not know what was actually observed, what is still unknown, or which next step is safe while the internet may be unreliable.
I built RACE to turn that anxious moment into a calm recovery lane anyone can understand.
What it does
Describe only what happened. RACE separates observations from assumptions, explains the likely scope in plain language, and gives three low-risk checks. It then opens a recovery workspace where an already-installed local Ollama model can propose one bounded next step even when cloud access is unavailable.
The proposal is never trusted blindly. RACE validates its structure and safety boundary, lets the person edit it, and requires review before exporting a Codex-ready packet. The packet gives Codex the evidence, unknowns, completed checks, and exact permission boundary needed for deeper help. RACE never claims to identify an attacker, never changes network settings, and never sends anything automatically.
How it works
- Capture the symptom in everyday language.
- Split what was observed from what remains unknown.
- Follow reversible checks before disruptive steps.
- Optionally ask a localhost-only model for one no-tools proposal.
- Validate and edit that proposal.
- Export a privacy-bounded case for Codex only after human review.
The full deterministic recovery path works with no dependency installation, API key, cloud account, or paid service. Ollama is optional and is never downloaded or started by RACE.
How I used Codex and GPT-5.6
GPT-5.6 through Codex helped transform a broader defensive-response concept into this relatable consumer product. Codex materially assisted the interaction design, threat and privacy boundaries, implementation, model-output validator, adversarial test generation, keyboard replay, accessibility QA, and the final real-motion demo. I made the product, safety, identity, claims, and release decisions.
Challenges
The challenge was being useful without pretending to know the cause. The same symptom can come from Wi-Fi, the service, the router, an ISP path, or intentional disruption. RACE therefore labels uncertainty, avoids attribution, and blocks dangerous or unsupported advice. Another challenge was keeping help available during an outage, which led to the optional local-model lane and the always-available deterministic fallback.
Accomplishments
- A working recovery flow that starts with a human description, not network jargon.
- Optional offline AI with no tools and a deterministic safety validator.
- Human-approved Codex packet with explicit runtime provenance.
- 18 focused tests, zero failures.
- Desktop and phone accessibility scans with zero reported violations or incomplete checks.
- Zero automatic network changes.
What I learned
The most reassuring tool is not the one that sounds most certain. It is the one that reduces uncertainty honestly, gives a safe next move, and makes escalation easier when deeper help is needed.
What's next
More outage patterns, platform-specific recovery guidance, signed incident receipts, accessible mobile packaging, and deeper Codex workflows that remain bounded by the user's evidence and approval.
Log in or sign up for Devpost to join the conversation.