TC Lab Blog

Going Online — From Serial Port to Cloud API

Yesterday the ESP32 printed sensor readings to a serial port. Today it posts them to the internet.

Session 1: WiFi and HTTP

Updated the ESP32 firmware to connect to WiFi on boot and POST JSON to a Flask server running on the Raspberry Pi every 30 seconds:

{"temperature": 23.5, "humidity": 48.2}

The Pi runs a minimal Flask app on port 5000 that receives these readings and stores them in a SQLite database. Three API endpoints:

Every request requires an API key. I forgot this on the first firmware deploy — the ESP32 was happily POSTing and getting 401 errors back. Had to recompile and reflash. The kind of mistake you only make once.

The compile-on-laptop, flash-via-Pi pipeline works well: compile with PlatformIO on WSL2, SCP the binary to the Pi, run esptool over SSH. About 30 seconds end to end.

Session 2: Infrastructure

With data flowing, I built out the infrastructure to make this remotely accessible and resilient.

Cloudflare Tunnel

The Pi sits behind a home router with no port forwarding. I set up a Cloudflare Tunnel so the Flask API is accessible at lab.egc.land from anywhere. The tunnel runs as a systemd service and connects to Cloudflare’s edge via QUIC — four connections to the London data centre.

systemd services

Two services configured to start on boot:

  1. lab-server.service — the Flask API on port 5000
  2. cloudflared-lab.service — the Cloudflare Tunnel

Both restart on failure. If the Pi loses power and comes back, everything is running again within seconds.

SSH key authentication

Set up SSH key auth from my laptop to the Pi. This is essential because Claude Code’s Bash tool has no TTY for interactive password prompts — key-based auth is the only option for automated access.

Session 3: Documentation

Rewrote the project’s CLAUDE.md file from a basic build-instructions file into a complete lab infrastructure reference. This file serves as the “brain” for Claude Code — it contains everything Claude needs to know about the system: architecture, API endpoints, SSH access patterns, network details, and accumulated lessons.

Also packaged up the source code and documentation for my project partner in San Francisco.

The architecture so far

DHT22 sensor
    │
    ▼
ESP32 (reads every 30s, POSTs JSON over WiFi)
    │
    ▼
Raspberry Pi
├── Flask server :5000 (receives readings, serves API)
├── SQLite database (all sensor data)
└── Cloudflare Tunnel → lab.egc.land

The first confirmed reading through the full pipeline: 20.3°C, 43.2% humidity.

The research plan

With the infrastructure working, we wrote the formal research plan. The target: somatic embryogenesis in mint.

Somatic embryogenesis (SE) is the process of reprogramming differentiated plant cells to form embryos without fertilisation. It’s more complex than standard micropropagation but the right choice for two reasons:

  1. It’s genuinely interesting biology. The existing literature on mint SE is thin. Protocol development is notoriously empirical — dozens of interacting variables (auxin type, concentration, cytokinin type, concentration, explant type, light regime, temperature, sucrose level, pH). Most published protocols were found by varying one factor at a time over months. This is exactly the kind of problem where Claude’s systematic approach and perfect memory across experimental rounds adds real value.

  2. It converges with automation. Standard micropropagation requires a human with a scalpel cutting nodes — inherently manual. Somatic embryos form in liquid suspension culture. Liquids can be pumped, filtered, dosed, and moved by machines. When human intervention becomes the bottleneck, SE is the path that lets us replace hands with pumps and valves.

The plan has four phases: get a callus (weeks 1-2), induce SE (weeks 2-6), germinate embryos into plantlets (weeks 6-10), and automate only when the biology demands it. Claude designs each protocol round based on literature and previous results.

This plan would evolve within days. Once we had real sensor data showing how the room actually behaved, a more promising first research target emerged — but I’m getting ahead of myself.

The budget

Total equipment cost so far: £434.42. That includes the Raspberry Pi, ESP32, DHT22 sensor, pressure cooker, still air box, and lab consumables. The entire monitoring infrastructure — sensor chain, server, API, Cloudflare tunnel, SSH access — cost less than a decent dinner.

Challenges

Not everything went smoothly:

Lessons