Team, Award & Field Projects | eecczz's page

Smart CCTV Fall Detection

Thu, 13 Aug 2026 00:00:00 +0000

BACKEND · AI · SYSTEM CASE STUDY

An Edge-AI monitoring project that reduces nursing-home RTSP false positives through three-stage pose, fall-object, and site-context zero-shot validation.

Role Edge-AI field internship · data, models, post-processing, field validationValidation Live nursing-home RTSP environment

Key Screens

Troubleshooting

1. Pose estimation alone was unstable for occlusion, distance, and vertical falls

Joint confidence, body angle, aspect ratio, and dwell-time heuristics changed with camera angle and occlusion. Pose remained the first candidate stage, while a fall/person detector became the second verifier.

2. The detector still confused sitting, bending, and lower-body occlusion with falls

Hard negatives for sitting, bending, and partial occlusion were added. Overlapping or adjacent boxes were grouped, and a fall survived only when confidence and body coverage were sufficient.

3. Model output alone could not interpret the site context

The final verifier compares candidate visual features with site-specific prompts such as a person fully lying on the floor, sitting on a chair, bending, or partially hidden by furniture.

System Flow

Pose keypoints and body tilt produce a fall candidate from RTSP frames.
A YOLO fall/person model verifies the cropped candidate.
Nearby boxes are grouped and confidence and coverage are measured.
Visual features and site-specific prompts perform zero-shot validation.
Only events above 0.70 confidence and the time condition are confirmed and sent to snapshots, DB, and alerts.

Gone

Wed, 01 Jul 2026 00:00:00 +0000

A choice-based story action-adventure game created as a four-person Coala club team project.

Tech focus: implementation, interaction design, backend/API structure, and project delivery
Repository or reference: portfolio project

Key Implementation Points

Choice-based Story

Provides different endings through choices in a ruined city attacked by AI.

Vertical Play

Tested a mobile-friendly 1:1 layout with game view above and controls below.

Club Team Development

Split Fungus-based story flow and action progression across team members.

Soulslike Game

Wed, 01 Jul 2026 00:00:00 +0000

A four-person Unity action game project set in a colosseum, built for an introductory creative engineering design course.

Tech focus: implementation, interaction design, backend/API structure, and project delivery
Repository or reference: portfolio project

Key Implementation Points

Boss Sequence

Built soulslike tension with wide-range attacks and strong effects.

Dodge and Combo

Designed a combat flow where players read attack motions, dodge, and counter.

Colosseum Stage

Created a final-battle space where the player and boss face each other.

SpeakUp - AI Agent Communication Coach

Wed, 01 Jul 2026 00:00:00 +0000

An AI coaching service that analyzes voice, gaze, posture, fillers, and expression for presentation and interview practice.

Tech focus: implementation, interaction design, backend/API structure, and project delivery
Repository or reference: https://github.com/eecczz/speech-coach

Key Implementation Points

Real-time Coaching

Analyzes speaking speed, fillers, and silence to provide feedback during practice.

Session Dashboard

Creates situation-based sessions and stores practice flow.

AI Report

Generates improvement reports from transcripts and key attention segments.

TRIPICK - Tourism Course Verification Platform

Wed, 01 Jul 2026 00:00:00 +0000

A user-participation tourism platform that combines TourAPI, GPS route data, reviews, and verification to recommend trustworthy courses.

Tech focus: implementation, interaction design, backend/API structure, and project delivery
Repository or reference: https://github.com/eecczz/tripick

Key Implementation Points

Verified Course Ranking

Calculates tourism course reliability from GPS route data, reviews, and verification.

Trust Score

Designed scoring beyond simple ratings by reflecting completion and verification data.

Tourism Data Use

Connects TourAPI and user-participation data so recommendations can be re-verified.

Jeonbuk Provincial Voice Chatbot

Mon, 01 Jun 2026 00:00:00 +0000

BACKEND · AI · SYSTEM CASE STUDY

A public voice-information service that retrieves Jeonbuk Province and municipal data with GraphRAG and delivers answers through STT and TTS.

Role Capstone · FastAPI, voice pipeline, GraphRAG, crawler operationValidation Capstone Grand Prize · Jeonbuk Governor's Citation

Key Screens

Troubleshooting

1. A new utterance had to interrupt an answer already being spoken

Instead of finishing STT → answer → TTS in one request, I rebuilt the pipeline around a real-time socket connection. When a new utterance is detected, the active response and playback are stopped before the next turn begins.

2. Daily crawling was slow, skipped pages, and overloaded the embedding server

I standardized dynamic-page extraction, content cleanup, and asynchronous crawling with Crawl4AI. Incremental, full, and per-site runs isolate failures and reduce BGE-M3 embedding load.

3. Slow RAG answers were not only a pipeline problem

I compared the same queries across different Qwen 3.5 model sizes, confirmed the model-selection impact on end-to-end latency, and changed the response model.

4. STT often missed the first syllable

VAD onset delay and confidence thresholds were tuned so the audio buffer reaches STT immediately after speech is detected.

5. Municipality URLs and SPA structures caused different collection failures

Base paths are configured per site, while normal HTML and browser-rendered SPA pages use separate collection paths.

Technology Choices

Technology	Why it was used
FastAPI	To compose asynchronous STT, retrieval, LLM, and TTS stages with different latency characteristics.
STT/TTS	To make public policy and civil-service information accessible through natural voice queries.
GraphRAG	To retrieve answers with document relationships and evidence preserved.
Incremental crawler	To refresh normal and SPA sites selectively rather than re-crawling everything.

System Flow

STT converts the user’s voice to text.
FastAPI receives the normalized question and conversation context.
GraphRAG retrieves relevant provincial and municipal data.
The LLM produces an evidence-grounded answer.
TTS returns speech synchronized with the character UI.

Next Implementation Plan

Show a source document and last-crawled time with every answer.
Add browser-pool, timeout, retry, latency, and failure-rate monitoring by pipeline stage.