Other Projects

Rather than skim the field, each module isolates one capability and builds something working around it — a decoder you can feel, a typed API client, an agent that actually calls tools, a diffusion canvas, a local model, a fine-tuned checkpoint. The modules stack into a single arc: understand generation, drive it through APIs, give it tools, give it other modalities, run it locally, then train your own.

The back half is the part I care about most: an end-to-end loop that turns public-domain text into a synthetic training set, fine-tunes a small model on it with parameter-efficient methods, and publishes the result — the same shape a real product team would use, done solo at small scale.

The opening module loads GPT-2 Medium from HuggingFace Transformers and builds an autocomplete around it, then makes the decoding strategy tunable — greedy versus sampling with temperature, top-k and top-p — so the effect of each knob on coherence and diversity is something you watch happen rather than read about.

Running the same prompt through every setting (and across English and Korean continuations) turns the abstract sampling parameters into intuition: where greedy collapses into loops, where temperature tips into noise, and where nucleus sampling lands in between.

The final module steps back from the code to a written analysis of automation bias and human oversight — why people over-trust automated output under stress, and how human-in-the-loop versus human-on-the-loop architectures change that. It is a deliberate counterweight to the rest of the work: a reminder that knowing how to build these systems is only half of knowing when and how they should be used.