S.01baseline
Sober Baseline
User completes reaction, balance, and memory tests while sober to set a personal reference — not a population BAC average.
[001]·jul_2026v0.2
BuzzBuddy
Hackathon prototype that detects performance deviation from a personal baseline. It does not determine sobriety or fitness to drive.
iOS app that estimates impairment by measuring deviation from your own sober baseline — reaction time, balance, memory — using an agentic AI examiner, not a BAC guess.
Award3rd · MLH × DigitalOcean AI Hackathon ’26
BaselineReaction / balance / memory, personal — not BAC
Toolsretrieve_baseline / request_test / analyze_deviation / update_confidence
ModelClaude 4.6 Sonnet · tool-calling loop
OutputDeviation confidence, never "safe to drive"
What I Contributed
- Built the FastAPI backend and the agentic tool-calling loop (retrieve_baseline → analyze_deviation → update_confidence → request_test) against DigitalOcean Serverless Inference.
- Designed the SQLAlchemy schema and Alembic migrations.
- Wired the AI-driven check-in flow into the iOS app's shared state, replacing a local placeholder test shuffle.
- Built, then removed before submission: a Twilio SMS designated-driver alert and a secondary agent for read-only driver Q&A.
- Tuned the examiner's system prompt (tone, benefit-of-the-doubt calibration on ambiguous evidence).
- UI passes on the check-in and verdict screens.
How It Works
4 stagesS.02test
Adaptive Testing
On check-in, the AI examiner requests specific tests one at a time based on what it still needs to know.
S.03analyze
Deviation Analysis
Each result is compared against the personal baseline, and a tool-calling loop updates a running confidence score with reasoning attached.
S.04verdict
Final Summary
The examiner concludes the session with a plain-language summary — it never tells the user it’s "safe to drive."
Evaluation & Evidence
- No automated test suite exists for this prototype (verified: zero test files in the backend).
- Verified in code: each verdict is grounded in a specific per-test percent-deviation from the stored baseline value, not a single opaque score.
- Verification during the hackathon was manual/demo-driven — no measured pass rates, scripted-scenario counts, or response-time benchmarks are available to report.
Constraints & Limitations
- Not medically validated and not more reliable than BAC testing — it measures deviation from a personal baseline only.
- No automated tests or regression suite.
- A Twilio SMS alert to a designated driver, and a secondary driver-Q&A agent, were built and worked during the hackathon but were removed before final submission — the current backend always ends a session with a plain-language summary rather than escalating to a contact.
- Several iOS tabs (Home/Events/Settings) are placeholder stubs not wired to real state beyond the check-in flow itself.
What I'd Improve Next
- Add an automated/scripted-scenario test suite.
- Decide whether to re-add and harden the designated-driver notification path, or remove its remaining UI references.
- Wire the remaining placeholder tabs to real state.