Jason Lernerman's Portfolio

MidnightSky is a startup game company (2 developers, 3 administrative staff) that runs a Minecraft Server with ~75 concurrent players (at all times of the day) and several thousand monthly active users where I lead feature development from July to October 2025.

The tech stack was a large (50K LOC+) Kotlin and Java project that stored data in a MongoDB database and eventually PostHog.

I designed several large new features from the design stage all the way to the implementation stage along with support from administrative staff. Then, we launched a new "season" (3 month period in which everyone starts the game from scratch with whatever new features came in the new update).

Funny that you ask, yes I in fact did. Once we released a new season of the game, I quickly learned we had zero analytics about how users were interacting with both existing features, and the new features I had created. This made it hard to know whether any of the thousands of magic numbers the game had needed to be tweaked.

I designed and implemented an interface for events to be sent from Kotlin to PostHog to allow for data-based decision making for all of the game staff. This addition started from neither me nor anybody else on the team knowing anything about analytics, to the first event, to the first graph, to leading a seminar amongst the team how to make more graphs, to eventually having the ability for any member of the team being able to make a graph for what they were looking to monitor without even needing to loop in a developer. All in all, a success much larger than I could have imagined at the outset.

Dataset AI is a startup AI company (3 engineers) where I worked as a Software Engineer from January to May 2025.

The tech stack was a (5K LOC+) React frontend, a (10K LOC+) Python backend and data stored within Google Cloud's Firebase database.

I collaborated with a team of 3 to build a web app which supported an excel-like editor which allowed for the user to speak to multiple AI model conversations in parallel. The main use case for this was comparing a fine-tuned model with a base model across many different prompts at once.

Yeah, I also learned how and then optimized a React app that had gotten slow using industry tools like react-scan, resulting in bringing the experience back to usable for the webapp.

No, later into this position I was reassigned to work on a different team within the parent company which maintained a (10K LOC+) React Native AI chat bot supporting Web, iOS, and Android for about 1 million monthly active users

Vue Technology LLC is the company that sponsors Vue development (and works closely with ByteDance, the company which owns TikTok) where I did a Software Engineer Internship from December 2022 to January 2023.

The tech stack that I worked on was an open source Rust project.

I implemented an algorithm called a call-flow graph to an open source project called oxc which enabled saving significant bandwidth when communicating between website and client code. These improvements were contributed to a Rust application that has 2 million weekly downloads on npm, it's called oxlint .

Yeah it was, you can take a look at some of the work in the PR here Open in File Explorer, the work was later merged within this PR Open in File Explorer.

toplines is a CLI, that I built in April 2026, that finds the files with the most lines in your project. It respects .gitignore and skips binary/asset files by default, then gives you a ranked list of the biggest files.

TypeScript running on Bun. It uses git ls-files under the hood to get the file list while respecting .gitignore, then counts lines concurrently across all files.

Check it out on GitHub Open in File Explorer.

promptlog is a Claude Code plugin I built that logs every prompt I send to Claude into a local file, so I can search my own history later.

I kept wanting to go back and find something I'd asked Claude Code a few days prior, and there was no built-in way to do it. Turned out to be fixable with a ~15-line shell script once I understood Claude Code's hook system.

Claude Code fires a UserPromptSubmit hook every time you send a message. The hook receives JSON on stdin with the prompt, session ID, and working directory. promptlog's hook script parses that with jq and appends each prompt to a timestamped log file under ~/.promptlog/{project_name}/, one folder per project.

Bash and jq. That's it, no runtime dependencies beyond what most dev machines already have.

It's published as a Claude Code plugin. Inside Claude Code, run /plugin marketplace add u9g/promptlog and then /plugin install promptlog@u9g. The source is available on GitHub Open in File Explorer.

Uno Royale is a single-player UNO card game where you compete against three AI-controlled opponents. It implements UNO the way I like to play it, along with the standard UNO rules that everyone plays with.

I built it in March 2026.

Vue 3 with TypeScript and Vite for the web app, Vitest for unit testing the game mechanics, and Capacitor to wrap it as a native iOS app.

Yeah, I added a "Make Computer Players Instant" toggle that makes the game feel fresh and like no other UNO game. In addition, I also built a screenshot pipeline using Playwright that automatically generates App Store screenshots across 5 device configurations (iPhone, iPad, and desktop in multiple orientations) by injecting crafted game states. The screenshots are regenerated on every commit via a pre-commit hook that also runs ESLint, Stylelint, Vitest, and ShellCheck all in parallel — so App Store screenshots are always in sync with the code, and if they change they get automatically staged into the commit.

You can check out the code on GitHub Open in File Explorer or download it on the iOS App Store .

LearnTensors is a web-based coding practice platform, that I built in March 2026, for tensor and matrix mathematics using Python and PyTorch. Think LeetCode but focused on tensor operations — you get a problem description, write your solution in a Monaco editor with real-time Python type checking, and run it against test cases. Try it at learntensors.com .

Vue 3 with TypeScript, server-side rendered on Cloudflare Pages with a D1 (SQLite) database. The editor uses Monaco with a WASM-compiled Python type checker (ty from Astral's Ruff project) for real-time diagnostics, hover info, completions, and inlay hints. GitHub OAuth for authentication.

Integrating ty as a WASM module to provide a full Python language server experience directly in the browser was the big one — it gives real-time type errors, signature help, and semantic tokens without any backend round-trips. I also extracted the layout engine into its own npm package (reactive-layout ) when I realized it was general-purpose enough to reuse.

Yeah, check it out on GitHub Open in File Explorer.

reactive-layout is a Vue 3 layout engine for building split-pane, tabbed interfaces with drag-and-drop. Think VS Code's panel system — resizable splits, draggable tabs, and drop-to-split — as a reusable package. It's published on npm .

I built it in March 2026 alongside LearnTensors, a coding practice platform I was building, when I realized the layout engine was general-purpose enough to be its own package.

Pure Vue 3 with TypeScript — zero dependencies beyond Vue itself. It uses Vue's reactivity system for the layout tree, provide/inject for extensibility, and the HTML5 Drag and Drop API for tab movement.

Recursive split panes that nest as deep as you want, drag tabs between panels or reorder within a panel, drop on panel edges to split horizontally or vertically, resizable panes with minimum size constraints, optional localStorage persistence, and SSR safety. The consumer provides their own panel content and tab icons via provide/inject, so the layout engine stays fully generic.

Yeah, there's a live demo and the code is on GitHub Open in File Explorer.

ColorPicker is a lightweight macOS menu bar app for picking colors from anywhere on your screen. Click the eyedropper icon, hover over any pixel, and click to copy the hex code to your clipboard.

I built it in March 2026.

It's a native macOS app written entirely in Swift using AppKit and Core Graphics. It runs as a menu bar utility (LSUIElement) with no dock icon — just a small eyedropper in the status bar.

The main challenge was color accuracy. On modern Macs with Display P3 screens, the standard NSBitmapImageRep API returns colors in the wrong color space, producing incorrect hex values. I solved this by reading raw pixel data directly from the CGImage, bypassing the color management layer entirely to get accurate sRGB hex codes.

Check it out on GitHub Open in File Explorer or download it on the App Store .

It's the website you're looking at right now — jasonlernerman.com . A portfolio site with a conversation-style UI that presents my work experience and projects as Q&A threads.

I built it in March 2026.

Vue 3 with TypeScript and Vite. It uses hash-based routing, a collapsible sidebar, and all the content is driven by a single JSON data file. Styling is a custom dark theme with the Lora serif font.

Yeah, check it out on GitHub Open in File Explorer.

Cerium is an AI chat application I worked on to learn AI model API's and the data structure of a typical AI chat app.

The frontend was a SwiftUI iOS native app. The backend was a TypeScript API through Convex (a serverless function provider similar to AWS Lambda or Cloudflare Workers). The database was Convex's builtin document store (like MongoDB).

I designed and implemented a SwiftUI mobile app similar to the ChatGPT mobile app.

Yeah absolutely, take a look at it here on the App Store .