Hrishi Olickel's Approach to Building Practical AI Tools

Table of content

Hrishi Olickel runs Southbridge AI and builds open-source tools that solve real problems he encounters while working with language models. His projects have a common thread: they do one thing well, ship fast, and come with honest documentation about their limitations.

The “Future Self” Documentation Method

Olickel titles his major guides “Everything I’ll forget about X” — a framing that sets expectations correctly. These aren’t theoretical overviews. They’re reference documents he writes for his own use, then publishes because others find them helpful.

His prompting guide opens with a clear statement: “This isn’t written by ChatGPT. Claude and friends had no hand making this guide.” The 2022 piece covers template differences between models, few-shot learning techniques, and complexity reduction — practical knowledge that ages better than framework-specific tutorials.

The evals guide from 2024 addresses a gap he noticed: everyone talks about why evals matter, but nobody explains how to actually build them. He breaks down the core problem:

How do you get a dumber intelligence to verify the output of a smarter intelligence, such that the results are a statistically stable proxy of success?

His answer involves simpler datasets (one entity per row, not batched), avoiding premade eval prompts from platforms, and starting with Python loops before reaching for tools.

Notable Open Source Projects

Lumentis (1.7k stars) generates documentation from meeting transcripts and unstructured text. Run npx lumentis and it produces a full documentation site. The tool exists because Olickel needed to turn long-form content into something people would actually read.

Mandark is a minimal AI coding assistant at around 80kb. The tagline captures its philosophy: “Simple AI coder that can do most of my work for me, including working on himself.” It handles file editing, package installation, and cost estimation before execution.

Wishful Search enables natural language queries against JSON arrays. Instead of writing filtering logic, you describe what you want. The library handles the translation.

Rakis is a browser-based inference network — AI models running directly in browsers that form a distributed network. He published analytics from a month-long stability test, showing the messy reality of running such a system.

From Maritime Tech to AI

Before Southbridge, Olickel built AI systems at Greywing, a maritime logistics company. His posts about one-click crew changes show how he thinks about AI in production: corrective workflows where humans verify AI outputs rather than blindly accepting them.

The maritime work also produced his Better RAG series, which argues against treating RAG as a solved problem. His approach emphasizes using the full document (“Use the whole brain, luke”) and iterative retrieval rather than single-shot queries.

Personal AI Workflows

Olickel documented solving carpal tunnel with Whisper and GPT-4, showing how he chains speech-to-text with an LLM to write code by voice. The post includes audio samples, exact commands, and an honest assessment of where the approach breaks down.

His writing pipeline describes using AI for editing and structure, not generation — a distinction he considers important for maintaining voice and original thought.

Research Background

His academic work includes formal verification for smart contracts (presented at ACM CCS 2016, code later joined the Ethereum Foundation), sensory substitution research, and DNA sequencing analysis. The research background shows in his technical posts: he cites sources, quantifies claims, and acknowledges uncertainty.

Where to Start