karim@portfolio ~ % cat about.md
I am a self-taught builder who works at the edge of agentic infrastructure and DeFi. My journey began in 2021, when I began experimenting with DeFi protocols and hunting airdrops, gaining a deep understanding of the underlying technology. In mid-2025, I decided to take a step back from Web3 and make my entry into Web2 for the first time. I began experimenting with agentic development, and what began as pure interest and fun, later turned into a routine, a workflow, a state of mind.
I currently build full-stack products solo, everything from Rust and Tauri desktop tools to on-chain trading systems, and I also handle business development in person across the San Francisco tech scene.
Everything I've worked on has been something I started and drove myself. I'm ambitious and I move on my own initiative. Even when I'm walking into something I don't fully understand yet, I figure it out and come back with results that hold up. That's how I believe true success is achieved - through consistent and efficient iteration.
- based
- San Francisco
- focus
- Agentic infrastructure / Business development
karim@portfolio ~ % ls ~/projects # 8 entries
A macOS desktop app that tails the transcripts your local coding agents already write (Claude Code, Codex) and renders every session as a live agent on a 3D radar, so your whole fleet is on one screen. Each globe heats from a dim ember to white-hot as its context window fills, subagents orbit their parent, and color marks the harness. Read-only and fully local: it never writes to your projects, makes no network calls, and needs no API keys. Tauri v2 with a Rust core and a React Three Fiber scene.
- rust
- tauri
- react-three-fiber
karim@portfolio ~ % open news.ycombinator.com/front # front page today
Hang on, this is one person? The Rust daemon, the transcript parser, the live subagent tree, the R3F radar, and not one of them feels phoned in. This does not read like a first project.
Solo, yeah. I came out of DeFi, a few years on-chain writing trading systems. The watching-for-drift instinct is just habit from running bots you cannot afford to have quietly go wrong.
I lead a small team building agent infra, and the instinct here, catching a run going wrong before it burns you, is the exact thing we have interviewed dozens of people for and not found. A little surprised you are still independent. Email is in my profile.
Appreciate it, genuinely. Sending you a note now.
Seconded. You can teach someone a stack in a month, you cannot teach the reflex to build read-only from day one because they have already been burned by automation that lied to them. That part is earned.
Went down a rabbit hole through your other repos after this. A native crypto terminal that signs its own Hyperliquid orders locally, a self-hosted news filter that scores your feed with an LLM, a caption editor and a teleprompter that both run on-device, all this year and all on top of Warden. Any one of them is a respectable solo project. The throughput at this quality is the part I cannot square with one person. Who is backing this, or is the honest answer nobody yet?
Nobody, honestly. They started as things I wanted and could not find, so they were getting built either way.
Ran this next to a Claude Code session while I worked on something else. Glanced over and the agent's globe was stuck pulsing in place, re-reading the same three files in a loop, a good ten minutes before I would have caught it myself. Would not leave a tool that can write running unattended on a client repo, so the read-only stance sold me by itself.
That loop is the exact thing that made me build it. Mine used to stall quietly and I would find out twenty minutes and a lot of tokens later.
The consistency is the tell for me. I self-hosted his dev-signal filter last month, bring-your-own-key, scores your feed with a small model and runs for a couple dollars, and it has the same fingerprint as Warden: local by default, nothing phoning home, built like he has to live with it himself. You do not fake that across five projects. That through-line is the part I would not know how to hire for.
I run engineering at a dev-tools company and I read this thread twice to make sure I was not missing the catch. One person, and the pieces fit together the way they usually only do after a team has argued about them for a year. People who build like this do not stay independent for long.
Native Rust, not another Electron thing squatting on a gig of RAM in the background. Rare to see someone skip the easy cross-platform route on a solo build.
How are you parsing the transcripts? Claude Code's on-disk format has shifted under me twice and broke a script I had. Pinned to a version, or tracking a schema?
Thin adapter per agent, one for Claude Code and one for Codex, so a format change is a small patch instead of a rewrite. Codex moved last month and it was about ten lines.
Went in skeptical, most agent observability is dashboards you open once and forget. A live radar you actually glance at is a different thing, it earns the look instead of demanding it. I have watched funded teams ship a worse read on this exact problem.
Slightly off topic, but a live 3D radar of your agents is very mission-control as a vibe. Is the full constellation always spinning in your face, or can you keep it low key? At 2am I am not sure my eyes want the whole animated starfield.
It calms down on its own. Motion is status, so idle agents just breathe slow and dim and a quiet night is a quiet radar. Only a live run lights the thing up.
Does any of this leave the machine? I want to point this at work repos and cannot have transcripts phoning home.
Nothing leaves the box. No network calls, no API keys, it only reads the logs your agents already write to disk. The read-only, offline design already assumes it should not be trusted with your network.