He Built the Diagram He Was Proud Of. Users Glanced at It Once and Left.

The first version of backthread.dev put a live architecture map dead centre, the thing Jevgeni was proudest of building. Then he watched people use it. They looked at the map once, and left. What they kept coming back for was the why: the decisions, the trade-offs, the risks their coding agents had quietly accepted on their behalf. So he flipped the entire app around. The diagram became one of five destinations. The decisions moved to the middle.

How It Started

Jevgeni had been building projects with coding agents and kept hitting the same wall. Give away too much control and you lose the thread and when the agent eventually gets stuck, you can't help it, because you no longer understand how anything works under the hood.

It compounds. The more agents ship, the worse the gap gets. To actually understand a system you have to either sit inside the thinking as it happens, or reverse-engineer it later and neither is cheap. Meanwhile developer tooling was exploding in every direction except this one. Nothing existed to help a developer and their agents keep a working mental model of the system without overspending on it.

That gap is backthread.dev.

Building From the Baltics

Jevgeni's Baltic advantage is unglamorous and enormously practical: low admin. Setting up the company, the bank account, and the accounting in Estonia is all digital and genuinely easy. The bureaucratic drag that quietly kills solo founders elsewhere simply isn't there.

That matters more than it sounds, because it enables the experiment he's actually running. Building backthread.dev solo lets him test how far a one-person, fully agentic company can go before he hires or finds a partner. In a country where incorporation is a form rather than a project, that experiment is viable. Elsewhere, the admin alone would force a second pair of hands.

How It Works

A CTO, tech lead, or engineer signs up with GitHub, connects their repo and their coding agent, and immediately sees the biggest risks backthread.dev found inside their agent's session transcripts. From there they can browse the system and see what the agent decided for each module, which trade-offs it quietly accepted, and which assumptions it made. As the team keeps building, new decisions and risks surface automatically.

The product got there in three months and three distinct shapes: it started as a simple system diagram of a codebase, then gained time travel so you could watch the system change, then came the one nobody else touches, agent session transcripts. Together they produce a view of an AI-built system that doesn't currently exist anywhere else.

Under the hood
The app is Vite + React + TypeScript, with React Flow for the canvas and ELK for layout. Ingestion runs on Cloudflare Workers plus a Firecracker sandbox. The repo is cloned into an ephemeral sandbox, read statically without ever running your code, then the sandbox is destroyed, nothing is retained. On the local path, redaction happens on your machine before anything is sent, and that client-side code is open source (npx backthread) so you can audit it. Data lives in Supabase Postgres behind row-level security. The LLM work is cost-tiered: Gemini Flash-Lite for the bulk, Claude Sonnet only where it counts because, as Jevgeni puts it, "a tool that costs more than the engineer it helps is dead on arrival."

The Hard Part

The moment that changed the product wasn't a crisis, it was a rebuild he chose to do.

He'd been certain the diagram was the product. Watching real usage proved otherwise, and the fix wasn't a tweak; decisions had to become the centre of the app and the architecture map had to be demoted to one destination among several. That cost a genuine rebuild. The lesson is one most founders learn expensively: the thing you're proud of building is rarely the thing the customer actually wants.

The counterweight came from five tester interviews at the end of a design sprint. Jevgeni had deliberately avoided naming the problem, waiting to see what testers would call it and listening for whether they'd reach for his words or their own. They used their own. Hearing the pain described back to him in someone else's language was the moment it stopped being his personal itch and started looking like a problem a lot of people quietly have.

What's Next

The three-year ambition is unusually concrete: backthread.dev becomes a normal step in how software gets built, as standard as code review or CI.

Right now, keeping up your mental model of an AI-coded project isn't part of anyone's workflow. It just doesn't happen and the bill arrives weeks later when something breaks. Jevgeni wants that catch-up work to be called backthread.dev by name, the way git and grep became verbs.

For the product, that means moving from "read the decision log" to "the log answers you" in any language, for any agent. For the team, he stays solo as long as he can, pushing the fully agentic one-person company as far as it goes. And for the market, "I need to understand what my agents built" ends up with one obvious answer.

The Toolbox

The thing Jevgeni can't work without is context management. Getting the agent the right information without drowning it in noise. In practice, that's Claude Code as a harness, running a fleet of agents in parallel git worktrees, each taking a ticket end-to-end: implement, review, merge, deploy. Most of his own time goes into planning; implementation gets handed off. "It's the most leverage I've ever had as a mostly-solo builder."

It's also, he notes, exactly the workflow that creates the problem backthread.dev solves. The more the agents ship, the less of it lives in his head. The tool he relies on most is the reason he needs the product he's building.

His recommendation is refreshingly against type. Not a book, not a podcast as "the algorithms recommend better than I could." Instead: spend more time outdoors, alone, with no content at all. That's where his best ideas come from.

🧵 How you can help
Jevgeni is looking for design-partner teams, 2–15-person shops shipping most of their code through Claude Code, Cursor, or Codex, who feel the "nobody fully understands what we shipped last month" pain. He wants teams who will point backthread.dev at a real repo and tell him precisely where it's wrong. If that's your team, or you know one, email him directly: jb@backthread.dev