One algorithm, implemented twice — in TypeScript and embedded C — and stitched together by CI

Ornamental-fish breeders run 10 to 100 tanks in a single room. A failed heater or a leak becomes irreversible if you notice it a few hours late.

Measurement in this domain is awkward, though. Electrodes drift out of calibration as they age. Alert on absolute thresholds and every drift produces a false alarm, until eventually nobody reads the notifications at all.

Catching real anomalies without false alarms meant looking at the speed of change rather than the absolute value. That much is a straightforward design decision.

The hard part was that both the cloud and the device needed that judgement. A device with no connectivity still has to detect anomalies on its own; the cloud has to reason across many devices. So the same logic was required in two places — one in TypeScript, the other in C on an ESP32-C3.

I could see the future where those two implementations drift apart.

Write both, and stitch them together with a contract test.

The decision fits a least-squares regression over the samples in a window and takes slope × window width as the effective Δ — more robust to noise than an endpoint difference. That is implemented on both the TypeScript and the embedded-C side, and the C implementation's real output from deterministic fixed inputs is committed as golden fixtures, which the TypeScript tests then check themselves against. Change one side alone and CI always breaks.

Two teams' documents get checked against the same master copy, every time. Change one side only, and it comes straight back.

What I gave up was DRY. I accepted two copies of the same algorithm. Since the languages and runtimes cannot share code anyway, I decided what gets shared is the behaviour.

The embedded side runs the regression over a 16-point ring, accumulating in double to avoid loss of significance, and only fires after N consecutive threshold crossings.

For power loss, the flash ring buffer is built so that a half-written record fails its CRC and only that single record is discarded — the good records on either side always survive. Flash operations are injectable, so the same C runs against an in-memory fake in host tests, with no hardware.

Device auth is a three-party HMAC that deliberately keeps no nonce store: samples are idempotent on (device_id, seq, ts) and hello/config are read-only, so a replay is harmless. The reasoning for not having one is written down as part of the design.

The ingest path ACKs first and defers alert evaluation and counters until after the response.

30 commits between 2026-07-03 and 07-27. The firmware is 7,429 lines of C and headers on ESP-IDF 5.2.2. The database has 25 RLS policies, and the device-secret table alone carries no policy whatsoever, so not a single row is reachable through PostgREST — the plaintext exists only on the QR code in the box. The time series is monthly-partitioned, with 7 pg_cron jobs driving rollups, retention and partition maintenance.

CI has two unusual gates. copy-lint fails the build if marketing copy claims something the hardware cannot do, such as detecting ammonia; each of its 21 rules carries a written reason. bom-lint fails if the bill-of-materials arithmetic drifts more than 10% from the pricing model's COGS. The firmware build also has a size gate: it fails above 1,400 KB against a 1,600 KB OTA slot.

Sensors, gateway, time-series database, notifications, billing — and a CI that constrains what you are allowed to claim — carried end to end by one person. On projects that lift field sensors into the cloud (manufacturing, agriculture, facility maintenance, remote check-ins) there is no translation cost between layers.

In particular, the technique for keeping one piece of logic identical across two runtimes — golden-fixture contract tests — is not IoT-specific. It applies just as well to a calculation that must agree between mobile and server.

Work that spans hardware and cloud falls under SaaS / industry DX development. For vertical DX engagements, see the lumber-industry case study.