Vigil

Someone is always watching over them, so you don't have to worry.

AI eldercare guardian — a single camera that understands an elderly person's day, lets family ask questions in plain language, and proactively alerts when something is wrong.

Built for the ClickHouse × Hex hackathon.

Stack

• CV: YOLO + MediaPipe pose (fall/activity) + CLIP embeddings + Qwen-VL captions
• OLTP: Postgres (raw events)
• OLAP: ClickHouse (behavioral time-series, vector search, baseline/anomaly rollups)
• Agent: Claude (NL check-ins + alerts)
• Surface: Hex (family routine dashboard + anomaly view)

Grounded in real data

• Toyota Smarthome / ETRI-Activity3D — real elderly daily activities
• UR Fall Detection / Le2i — real recorded falls
• CASAS — real months-long in-home behavioral baselines

See scripts/get_datasets.md for download links + ingest order.

Architecture

camera / dataset video
  -> CV (vigil/cv.py): YOLO detect + MediaPipe pose (fall) + CLIP embed
  -> Postgres (OLTP, sql/postgres_schema.sql)        # raw events
  -> ClickHouse (OLAP, sql/clickhouse_schema.sql)    # time-series + vectors + baselines
  -> agent (vigil/agent.py): NL -> ClickHouse -> answer
  -> Hex (hex/README.md): family dashboard + anomaly view

Quickstart

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env            # fill in ClickHouse Cloud + Postgres + Anthropic

python scripts/setup_db.py      # create schemas (Postgres + ClickHouse)
python scripts/seed_baseline.py # 30-day baseline w/ a built-in anomaly for the demo

# grounded-in-real-data: ingest real dataset clips
python -m vigil.ingest --dir data/urfd --location living_room

# ask it anything (the agent)
python -m vigil.agent "Did mom eat breakfast today and how is she doing?"

# live demo: webcam -> fake a fall -> alert fires
python -m vigil.ingest --webcam

Layout

vigil/          core package (config, events, db, cv, ingest, alerts, agent)
sql/            postgres_schema.sql + clickhouse_schema.sql
scripts/        setup_db.py + seed_baseline.py + get_datasets.md
hex/            dashboard build notes + queries
DEMO_SCRIPT.md  the 2-min pitch that wins the room

Demo-day priority (build in this order; never cut 1-3)

1. Real dataset -> CV -> events -> ClickHouse (grounded in real data)
2. Fall detection -> alert fires (the emotional climax)
3. One NL query -> ClickHouse -> answer (the agent)
4. One anomaly view in Hex (the ambition)