Engineering notes

Systems that were built, and what they taught back

Each post below takes one machine learning system that was actually built: what the problem really was, how the system was put together, the background needed to follow the field, how the industry solves the same problem, and what stayed useful once everything had shipped. The focus is on principles and architecture rather than step-by-step instructions.

REC

Recommendation and semantic search for a content app

The four layers of a multimodal recommender: from raw content analysis to vector representations, hybrid retrieval, and multi-source ranking.

recommender systemssemantic searchembeddingscollaborative filteringANN
PAD

Face anti-spoofing: telling a real face from a replay

Four architectural approaches to presentation attack detection, a mechanism for separating style from content, and how the field standardized measurement.

anti-spoofingdomain generalizationthreshold calibrationISO/IEC 30107-3biometrics
LM

A Vietnamese language model: predicting the next word

Three tokenization paths for Vietnamese, five architectures built from scratch, and the condition under which two perplexity numbers can be compared at all.

language modelstokenizerperplexityLoRAvietnamese
DET

Vehicle detection and the road down to the edge

Detecting vehicles in bad weather on constrained hardware: two-tier augmentation, knowledge distillation, and the path down to the device.

object detectionknowledge distillationmAPedge deploymentaugmentation
TS

Environmental monitoring: data infrastructure before models

Storing data that has both a time and a space dimension, forecasting each parameter on its own, and how the field measures forecast error properly.

time seriesforecastinggeospatial dataARIMAMASE
META

Five principles that carry across every ML system

The engineering habits that hold their value in every domain: build checkable invariants, guard against leakage, set a frame of reference, choose the right measure, and keep the configuration with the artifact.

ml practicedata leakagebaselinesmetric selectiontechnical debt