1
CHAT
VIDEO
DOC
WEB
✓
ON-DEVICE SAFETY LAYER
LAVOCADO
Private, on-device AI that turns risky screen moments into a safer next action.
Windows + macOS • Local visual detection • Multi-frame verification • Supportive intervention
“Most blockers know the URL. LAVOCADO understands what is actually visible.”
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
2
THE PROBLEM
Risky content no longer lives on a list of websites.
Blocklists know where you are. They do not know what is actually visible.
S
SOCIAL
C
CHAT
V
VIDEO
F
FILES
ONE
SCREEN
CONTENT MOVES ACROSS APPS
1
URL BLIND SPOT
Feeds, attachments, video frames and local files bypass site lists.
2
SINGLE-FRAME NOISE
Brief or ambiguous imagery can trigger a brittle classifier.
3
CLOUD PRIVACY COST
Uploading the screen exposes the user’s most sensitive context.
WHO IT SERVES
Privacy-conscious people who want voluntary, cross-app digital wellbeing support.
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
3
THE SOLUTION
One local safety loop across the entire desktop.
Three specialized stages turn raw pixels into a calmer next action.
01
DETECT LOCALLY
Native capture + on-device visual models
PIXELS STAY HERE
02
VERIFY BEFORE INTERRUPTING
ROI recheck • rescue tiles • tracking • fresh frames
TRUST BEFORE ACTION
03
SUPPORT THE NEXT ACTION
Pause → Breathe → Ready → optional AI practice
AGENCY, NOT SHAME
✓
Screen pixels are never stored or sent to the LLM.
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
4
WORKING PRODUCT
Built as a usable desktop product—not a one-screen demo.
A complete journey from setup and live diagnostics to intervention and local history.
Dashboard overview
PROTECTION
Running
ON DEVICE
2
640
3
UI RECONSTRUCTION
Live diagnostics
LIVE SIGNAL
Capture native
Decision clear
Temporal 2 / 3
Track active
UI RECONSTRUCTION
Ready + AI companion
Choose your next action
AI companion
Take one slow breath…
Type here
READY
UI RECONSTRUCTION
WINDOWS + macOS
MULTI-MONITOR
APP + HOST RULES
LIVE EXPLAINABILITY
EN + 中文
OFFLINE AI FALLBACK
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
5
TECHNICAL ARCHITECTURE
Context policy and visual evidence stay separate by design.
Only NORMAL contexts enter Vision; explicit rules can block or bypass without changing pixel classification.
FOREGROUND CONTEXT
App ID + hostname only
NATIVE CAPTURE
DXGI / ScreenCaptureKit
CONTEXT POLICY
BLOCK
BYPASS
NORMAL
CHANGE
GATE
PREPROCESS
PRIMARY
DETECTOR
ROI / TILE
VERIFY
TRACK +
TEMPORAL
LAVOCADO SERVICE
Per-monitor state machine
INTERVENTION
Pause → Breathe → Ready
LOCAL SQLITE • minimal event metadata • rules • no pixels
FORCE_BLOCK > FULL_BYPASS > NORMAL
DASHBOARD → CHILD PROCESS
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
6
AI STACK
Specialized AI components cooperate without sharing authority.
Each model has one job—and one explicit limit.
01
PRIMARY VISION
NudeNet 640m
OR YOLO11 NSFW Small
Emits typed visual evidence:
label • confidence • bounding box
DOES NOT DECIDE THE PRODUCT ACTION.
02
REGION RANKING
Viddexa
Nano / Mini
Prioritizes high-risk tiles
under a limited compute budget
CAN NEVER BLOCK BY ITSELF.
03
GENERATIVE SUPPORT
OpenAI-compatible LLM
+ local fallback
Questions • five-round practice
• calm closing
NEVER RECEIVES PIXELS, URLS OR SCORES.
PINNED REVISIONS + FILE SIZE + SHA-256 + OFFLINE RUNTIME SELF-CHECK
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
7
PRECISION + EFFICIENCY
The system spends compute—and trust—carefully.
A candidate must survive spatial rechecks and temporal confirmation before interruption.
CHANGE
GATE
FULL-FRAME
EVIDENCE
ROI
RECHECK
RANKED
RESCUE TILE
TRACK SAME
REGION
2 HITS / 3
FRESH FRAMES
FRAME 101
candidate
HIT 1
FRAME 102
same region
HIT 2
INTERVENTION
Pause
Breathe
Ready
BLOCK
STATIC FRAMES
Skip expensive inference; periodic scans remain.
HIGH CHANGE
Temporarily switch to aggressive scan mode.
IDENTITY SAFE
Different monitors, tracks and duplicate frames never confirm each other.
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
8
PRIVACY + INTERVENTION
The most sensitive data never crosses the device boundary.
Local perception remains local; optional generative support is text-only and user initiated.
INSIDE THE DEVICE
✓
SCREEN FRAMES
memory only
ROI + TILES
memory only
LOCAL MODELS
on device
RULES
local
EVENT DB
metadata only
DIAGNOSTICS
allow-listed scalars
OPTIONAL • TEXT ONLY
Chosen LLM provider
today’s count
coarse time bucket
user-entered text
NO pixels • NO URL • NO confidence score
HUMAN-CENTERED INTERRUPTION
3s
PAUSE
6s
BREATHE
READY
CHOOSE / AI COMPANION
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
9
ENGINEERING EVIDENCE
A real cross-platform system with explicit architecture guardrails.
Typed boundaries, offline model integrity and automated tests make the prototype reviewable.
100
APPLICATION PYTHON MODULES
13.3K
APPLICATION LINES
58
TEST MODULES
471
TEST FUNCTIONS
4
PINNED MODEL PACKAGES
ARCHITECTURE GUARDRAILS
Context and Vision do not import each other
One typed decision boundary—not blocked: bool
Overlay cannot capture; Vision cannot create UI
Pinned revisions, file size, SHA-256 and offline load
CROSS-PLATFORM CI
✓
PY 3.12
Windows + macOS
✓
COMPILE
Windows + macOS
✓
CONTRACTS
Windows + macOS
✓
UNIT TESTS
Windows + macOS
✓
PACKAGE
Windows + macOS artifacts
2 NATIVE DESKTOP PLATFORMS
1 TYPED DECISION BOUNDARY
OFFLINE FROZEN SELF-CHECK
MODEL DIGEST VERIFICATION
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED
10
ROADMAP
From a working prototype to a trusted daily safety layer.
The core loop exists. The next milestone is evidence, distribution and user trust.
NOW
Cross-app local protection
Context rules + visual verification
Multi-monitor intervention
Explainable local dashboard
Optional privacy-limited AI companion
NEXT
1
VALIDATE
Hardware matrix + performance regression
2
MEASURE
Reproducible FPR, recall, latency and power
3
SHIP
Licenses, signing, notarization, installer
4
LEARN
Opt-in user studies + accessibility
Better AI should not only recognize risk.
It should preserve privacy, reduce uncertainty, and help people choose what happens next.
LAVOCADO / AI BUILDERS HACKATHON 2026
PRIVATE • LOCAL-FIRST • HUMAN-CENTERED