Josh Nalven

A CLI that runs prompts through Anthropic's Batch API for half the cost. It hides the submit-poll-retrieve lifecycle behind a few plain commands, so using batch pricing feels no harder than a normal prompt — and AI coding agents can offload expensive queries to the background and pull the answer back when it's ready.

50% cheaper inference · Submit-and-walk-away batching · Built for AI coding agents

I'm making a chatbot for my town on a volunteer basis to experiment with different production RAG strategies, and learn about them in practical terms. It currently summarizes the township's website, and I'll soon expand it further to include town laws and adjacent bodies like the Board of Ed. This is still very rough around the edges :)

RAG testing ground · 1100+ chunks · Civic tech

A résumé you can talk to. Instead of a RAG pipeline, it ships my whole career corpus into Haiku 4.5's context window on every call — career history deepened with transcribed self-interviews so it captures my voice, not just bullet points. It speaks about Josh in the third person and declines to guess when something isn't in the corpus.

Built on hours of self-interview · Whole-corpus context (no RAG) · Anthropic API-powered

My submission for the 2026 Hugging Face x Gradio Build Small hackathon. Fine-tuned a LoRA based on Qwen 3 8B, training it to output narrative JSON responses that drive the game engine for a D&D-inspired improvisational RPG. I learned a ton about model design, datasets, SFT, game design, and conversation design.

Learned about fine-tuning & dataset curation · Trained a working "dungeonmaster" model under 8B · Had a blast!

Tests whether a vision model reproduces the bouba/kiki effect — the cross-cultural tendency to match rounded shapes to 'bouba' and spiky ones to 'kiki.' Rather than train a classifier, it uses CLIP to score shapes against ~200 sensory anchor words like 'acidic,' 'staccato,' and 'buttery' across nine domains, judging a shape by how it feels, not just its outline.

CLIP-powered · ~200 anchor words across 9 sensory domains · Interactive Hugging Face demo

Makes text embeddings something you can see. Many models have a square number of dimensions, so a word's embedding can be rendered as a pixel grid — a visual 'fingerprint.' A heatmap and similarity matrix line words up side by side, revealing that semantically related words really do share a look. A learning toy, on purpose.

Each dimension as a pixel · Side-by-side semantic fingerprints · Interactive Hugging Face demo