our open source projects. things we build to scratch our itch — tools, libraries, and experiments

Context Viewer

Break a conversation log into its parts and visualise how each grows across turns in token count and relative weight.

Context Viewer takes a conversation log, breaks it down into constituent components — system prompts, tool calls, generated artefacts, user instructions — and visualises how each grows across turns in token count and relative weight. It uses AI-assisted semantic chunking to flag redundancy and irrelevance, and runs entirely in the browser with support for OpenAI conversation formats.

GitHub Github Docs Docs Playground Playground Research Paper Research Paper

Blog posts:

Play Video

Megasthenes

Ask plain-language questions about any GitHub or GitLab repository — public or private — and get sourced, structured answers.

Megasthenes lets you ask plain-language questions about any GitHub or GitLab repository and get sourced, structured answers pinned to a specific branch, tag, or commit. It supports any model via OpenRouter, Anthropic, or Google, runs tool execution in an isolated container, and ships with built-in OpenTelemetry tracing and an LLM-judge evaluation system to track answer quality over time.

GitHub Github Docs Docs

Blog posts:

import { Client } from "@nilenso/megasthenes";

const client = new Client();
const session = await client.connect({
  repo: { url: "https://github.com/owner/repo" },
  model: { provider: "anthropic", id: "claude-sonnet-4-6" },
  maxIterations: 20,
});

// Stream events as they arrive
const stream = session.ask("What frameworks does this project use?");
for await (const event of stream) {
  if (event.type === "text_delta") process.stdout.write(event.delta);
}

// Or wait for the complete result
const result = await session.ask("How are the tests structured?").result();
console.log(result.steps); // All steps: text, tool calls, thinking, etc.
console.log(result.usage); // Token usage across all iterations

await session.close();

GRPO Trainer

Toolkit to fine-tune LLMs

grpo-trainer is a toolkit for fine-tuning LLMs to generate high-quality OCaml code. It supports supervised fine-tuning (SFT), RLVR/GRPO training, evaluation, and monitoring with feedback from the OCaml compiler and test suite.

GitHub Github

Blog posts:

# Clone repository
git clone https://github.com/nilenso/grpo-trainer.git
cd grpo-trainer

# Enter nix shell
make shell

# Run the script
scripts/bootstrap.sh

# Start model server
llama-server -hf unsloth/Qwen2.5-Coder-1.5B-Instruct-GGUF:F16 -c 4096 -ngl -1

Bean

A spreadsheet programming environment built to make spreadsheets maintainable, auditable, and reliable.

Bean explores the spreadsheet as a programming environment from the ground up. It consists of grid, a formula language interpreter that enables reactive calculation with dynamic arrays.

GitHub Github Playground Playground YouTube Demo Website Website

Blog posts:

Bean

Goose

A background job processing library for Clojure with native support for RabbitMQ and Redis.

Goose supports enqueuing jobs for async execution, scheduled execution, batch processing, and cron jobs — with native support for both RabbitMQ and Redis as message brokers. Error handling and retries are built in, and the message broker and metrics backend are pluggable. A jobs management console provides visibility into queue state and job lifecycle. It is currently used in production across multiple companies.

GitHub Github Docs Docs YouTube Demo
Goose
×
×
Bean
×
Goose
Want to see more? Check out our github.