Nokast Projects

LocalizeAI reads your newsletters locally and generates a focused daily brief — top stories, key insights, and original links — without sending your data to the cloud. It's built for AI researchers, PMs, founders, and developers who want real signal without sacrificing privacy.

• ✓Runs locally using small language models
• ✓No subscriptions, fully open-source
• ✓Designed for people who actually use AI every day

Import a trace of your agent from LangSmith, Langfuse, OpenTelemetry, or a plain call log. JevControl finds the small routing, tool-choice, and continue/stop decisions inside it, lets you pick which ones to test, then reruns the task with a smaller model making just those choices.

• ✓Reads real agent traces — no changes to your production agent
• ✓You approve every decision before it’s tested
• ✓Compares latency, tokens, cost, and agreement side by side

An independent, open-source take on TypeSafe’s Jev and System One, built for local inference on NVIDIA DGX Spark. Instead of generating text, spark-s1 answers typed Choice, Score, and Boolean questions over a state with calibrated probabilities — built for routing, guardrails, and other fast agent decisions.

• ✓spark-s1-4b-v6: 92.9% held-out accuracy, ~75ms latency on DGX Spark
• ✓Ships jev-control-core and jev-control-es, the models JevControl uses
• ✓Fully open weights on Hugging Face

Reads a retrieved context and query together to catch indirect prompt injection and context poisoning that keyword filters miss — trained via a teacher → judge → student pipeline, entirely on a single DGX Spark box.

• ✓99.4% recall / 0.6% attack-success rate vs. a 68.8% / 31.2% regex baseline
• ✓37ms p50 latency — small enough to run inline
• ✓Open model weights on Hugging Face

A tool-enabled agent (CLI + Discord) that searches, reads sources, and writes up findings, with optional Notion and Google Drive integrations. Bring your own local or OpenAI-compatible model.

• ✓Local web search via SearxNG, with DuckDuckGo fallback
• ✓Works with LM Studio, Ollama, OpenRouter, or OpenAI-compatible endpoints
• ✓Runs via Docker Compose or a standalone CLI

A small, readable reference implementation that makes the core pieces of an agent loop visible — model calls, tool boundaries, state, and harness-owned verification. Meant to be read and modified, not hidden behind a framework.

• ✓Minimal teaching example: model + tool + verify loop
• ✓Runs against a local Ollama endpoint by default
• ✓A starting point for your own custom harness

A local CLI that sits between your clipboard and any AI assistant, swapping emails, API keys, SSNs, card numbers, and other sensitive data for labelled placeholders — regex-based by default, with optional NER for names and locations.

• ✓100% local — no accounts, no network calls
• ✓Detects keys and tokens for OpenAI, AWS, GitHub, Stripe, and more
• ✓Whole-file .env mode for config files

Benchmarks local LLMs on a single NVIDIA DGX Spark under real, concurrent load — context limits, latency, tool-calling reliability — and publishes a live leaderboard instead of vendor tokens/sec numbers.

• ✓15+ models benchmarked with a public leaderboard
• ✓Tests tool-calling reliability, not just raw speed
• ✓Full methodology and per-model write-ups published