Fully offline viewer for Twitter/X group-chat exports: fuzzy search, timeline, media gallery, stats, and a year-in-review Wrapped. No server, no build step, no dependencies.

Describe an analysis in plain English. STATlee generates runnable Python or R, moderates it, runs it in a Docker sandbox, and explains the results. Full web app: auth, billing, rate limiting, streaming output.

Job discovery, end to end: a scheduled cloud scraper feeds a two-stage LLM scorer and a desktop dashboard, and a LaTeX engine tailors the resume for each posting without fabricating a word.

Local-first RAG chatbot over the 36k-article Xeno Series Wiki. Hybrid dense + BM25 retrieval with reranking; every answer cites the wiki pages it came from. Evaluated against a 200-question gold set.

Paste a LeetCode problem, get study material in three depths, streamed as it generates. It drives the Claude CLI instead of an API key, and verifies every code sample in a Python sandbox before you see it.

Four-arm study: does pretraining a small T5 beat prompting or LoRA-adapting a bigger code LLM at fixing Java bugs? Scored on CodeBLEU and on whether the patch compiles and passes JUnit. Execution inverts the ranking. LoRA wins CodeBLEU at 0.854 but RAG fixes more real bugs, 35.8% of 201.