TV Watch Time Tracker

An automated ML pipeline that uses an existing home CCTV camera to measure how long my 4-year-old daughter actually watches TV each day, and pings me on Telegram every afternoon with the number.

A weekend project that turned a TAPO C200 motion-clip stream into a daily one-number screen-time report — with a self-improving model and a human-in-the-loop correction channel built into the daily Telegram message.

What it does

TAPO C200 (living room) ──► TAPO Care Cloud
                                  │  Lambda 1: downloader (11:30 IST)
                                  ▼
                          S3 clips/YYYY-MM-DD/...
                                  │  S3 event per clip — fans out in parallel
                                  ▼
                          Lambda 2: processor (container)
                          ffmpeg → adult-filter heuristic →
                          MobileNetV3-Small @ 70% threshold
                                  ▼
                          S3 results/YYYY-MM-DD/...
                                  │  Lambda 3: notifier (13:30 IST)
                                  ▼
                          Telegram: "📺 25m 30s · 175 clips"
                          + 5 lowest-confidence frames for review
                                  │  user replies "1, 3"
                                  ▼
                          Lambda 4: correction handler (API GW)
                          → saves false positives to training/corrections/
                                  │  Sunday 02:00 IST
                                  ▼
                          Processor re-runs as retrainer
                          → fine-tunes model → uploads new weights to S3

Stack

Layer	Tech
Camera	TP-Link TAPO C200 + TAPO Care cloud
Compute	AWS Lambda (4 functions, ap-south-1)
Heavy compute	Container Lambda (PyTorch + ffmpeg + OpenCV)
Model	MobileNetV3-Small, fine-tuned via transfer learning from ImageNet
Storage	S3 (clips, frames, results, model weights)
Secrets	SSM Parameter Store (TAPO creds, Telegram bot token)
Notifications	Telegram Bot API
Dashboard	Streamlit (operational metrics + confidence trends)
IaC	AWS SAM (CloudFormation)

The model in one paragraph

MobileNetV3-Small, ~2.5M parameters, fine-tuned from ImageNet weights on 603 frames extracted from 13 short clips (1 frame every 2 seconds, OpenCV). 80/20 train/val split, 20 epochs, weighted sampler for class balance, light augmentation (crop/flip/colour jitter). Live inference samples 1 frame every 10 seconds and only counts a frame as "watching" above a 0.70 confidence threshold. An adult-size heuristic skips frames where the largest moving object exceeds 40% of frame height — saves compute and avoids false positives from adults walking through.

→ Full plain-English walkthrough in ML_MODEL_EXPLAINER.md. → PM-framed problem/architecture/metrics doc in PM_INTERVIEW_DOC.md. → Substack narrative for general audience in SUBSTACK_POST.md.

The self-improvement loop

The notifier doesn't just send a number — it also sends the 5 lowest-confidence "watching" frames of the day and asks the user to flag any that are wrong. Replies hit an API Gateway webhook → Lambda 4 saves the flagged frames as training/corrections/not_watching/. Every Sunday at 02:00 IST the processor re-runs as a retrainer: loads the corrections, samples an equal number of confirmed positives, fine-tunes for 10 epochs at LR 1e-4, uploads new weights to S3. Next cold start picks up the new model automatically.

Repo layout

.
├── lambdas/
│   ├── downloader/             # TAPO Care → S3 (zip Lambda)
│   ├── processor/              # ffmpeg + MobileNet inference (container Lambda)
│   ├── notifier/               # daily Telegram digest + review prompt
│   └── correction_handler/     # Telegram webhook → corrections store
├── model/
│   ├── train.py                # cloud training entry point
│   ├── train_local.py          # local training + per-class accuracy report
│   └── (mobilenet_v3.pt is gitignored — lives in S3)
├── infra/                      # SAM templates (template.yaml + phase1/phase2)
├── scripts/
│   └── register_webhook.py     # one-shot Telegram webhook registration
├── dashboard.py                # Streamlit operational dashboard
└── tests/

Cost

Component	Monthly
TAPO Care	~₹150
AWS Lambda (4 functions, ~175 clips/day)	~₹0 (free tier)
AWS S3	~₹30
Telegram Bot	Free
Total	~₹180

Setup notes

• AWS account in ap-south-1 (Mumbai); TAPO server URL must match camera registration region (aps1 for India, not euw1).
• TAPO username/password and Telegram bot token live as SSM SecureString parameters under /tapo-tracker/... — never in source.
• Local AWS credentials and .streamlit/secrets.toml are gitignored; production runs use Lambda execution roles, not access keys.
• infra/template.yaml is the primary SAM template; template-phase1.yaml and template-phase2.yaml are the staged variants used during initial rollout.