5engineering notesRSS ↗

Notes from the build log.

Short write-ups distilled from the project READMEs: the decision that mattered, the number that proved it, and the pattern worth reusing.

  1. BRONZESILVERGOLD

    VeriDoc Studio2 min read

    Zero inference APIs: ML in the browser tab

    VeriLens, VeriDoc Studio and DocuForge run detection, OCR and forensics client-side. What it took to make “the document never leaves the device” true.

    · WebAssembly · ONNX · Forensics

  2. BRONZESILVERGOLD

    AeroSentry2 min read

    Agents that act need a gate that says no

    AeroSentry lets AI agents schedule, inspect and alert on drone fleets — and stops every high-risk action at a deterministic policy check and a human approval.

    · LangGraph · Human-in-the-loop · Evals

  3. BRONZESILVERGOLD

    VeriLens2 min read

    When INT8 was the wrong answer

    Quantising VeriLens cut the model payload by two thirds. Then the liveness model lost the ability to tell a face from a photo of one — so it stayed FP32.

    · ONNX Runtime Web · Quantisation · Computer vision

  4. BRONZESILVERGOLD

    Retail Lakehouse ETL2 min read

    Fail the run, not the row

    Retail Lakehouse ETL reads a million defective sales lines. Its quality gate quarantines bad rows with a reason and only stops the job when the share crosses a threshold.

    · PySpark · Data quality · Star schema

  5. BRONZESILVERGOLD

    LakeFlow2 min read

    One set of transforms, two paths

    LakeFlow’s streaming path and its nightly backfill run the same code, so one pytest suite proves both. That mattered more than the throughput number.

    · Spark Structured Streaming · CDC · Idempotency