Pinned Loading

1. iron-batchiron-batchPublic

   LLM serving primitives in Rust: paged KV allocation, continuous batching, streaming HTTP.

   Rust

2. lvm-pass-lenslvm-pass-lensPublic

   LoRA fine-tuned a 120B LLM to classify LLVM compiler pass interactions. 53.8% → 82.1% accuracy.

   Python

3. bit-floorbit-floorPublic

   Experimental 1-bit / ternary LLM inference in Rust — weights as XNOR+POPCNT, GPU buffers managed by Drop, no dequantization ever.

   Rust

4. Axon-SearchAxon-SearchPublic

   Hybrid semantic search engine combining BM25 + FAISS retrieval, fused via RRF and cross-encoder reranked, backed by an async politeness-aware crawler.

   Python

5. Fine-Tuning-Qwen3-8BFine-Tuning-Qwen3-8BPublic

   Fine-tuned Qwen3-8B on math — includes the data leak I found in my own benchmark and the GSM8K regression I caused and fixed.

   Python

6. sift-coresift-corePublic

   Hybrid search from scratch in Rust: LSM storage, BM25, flat vector index, RRF fusion. Zero deps.

   Rust