I'm a software engineer and co-founder of Avartana Labs LLP. I build local-first systems, tooling for AI coding agents, and product interfaces that hold up under real use.

I'd rather publish a caveat than a number I can't reproduce.

Selected work  ·  How I work  ·  Stack  ·  Avartana Labs  ·  All repositories

Three things, in roughly this order of depth:

Agent and developer tooling. Resident daemons, MCP servers, and editor/agent extensions — the plumbing that makes a coding agent fast and predictable instead of clever and unreliable. Mostly Rust for the hot paths, TypeScript for the surfaces the harnesses actually speak.

Local-first data and retrieval. Keeping the human-readable source of truth canonical and the index disposable. Search that answers from local storage, degrades visibly when it can't, and fails closed rather than quietly shipping private text to a remote model.

Product surfaces. Cross-platform mobile and web interfaces — Flutter, SwiftUI, Next.js — written to ship, not to demo.

Resident filesystem search for coding agents.

Agents lose real time spawning find, rg, and grep -rn subprocesses. This keeps a resident path + trigram index behind a UNIX socket and answers those queries in-process, scoped to the workspace. Ships as a Pi Agent extension, a CLI, and a stdio MCP server.

Rust daemon · TypeScript · MCP · AGPL-3.0

Its README opens with a note that earlier versions quoted best-case warm-cache numbers as if they were session-level figures, and claimed a CI that did not exist — and the current figures are each labelled with their measurement scope. The open issues are mostly measurement-integrity findings I filed against myself: an integrity pin that covers 15 of 41 measurement files, a baseline suite that moved when re-run. That audit trail is the part I'd point a reviewer at.

Tern terminal integration for the pi coding agent.

Adds rich surfaces to a terminal agent without patching it: MCP-style handshake, native Mermaid rendering, WKWebView driving, pane capture, session mirroring. The extension deliberately never writes TSP frames — pi's renderer uses synchronized multi-write output transactions, and a second writer tears them.

TypeScript · MIT · published on npm

Every feature is labelled stable or experimental in a status table, and the native-rendering path is documented as the special case rather than the default — it runs stock pi in a shell block, outside Tern, or under -p. 136 tests pass on node --test; tsc --noEmit is clean.

A local-first memory engine for coding agents.

Markdown stays the canonical, human-editable source; SQLite/FTS5 is a disposable index that can be rebuilt at any time. Rust engine, TypeScript extension, BM25 lexical retrieval, and a fail-closed DLP layer that redacts credential classes before an excerpt can leave the machine.

Rust · TypeScript · SQLite FTS5 · AGPL-3.0

The interesting part is not the latency table. It's bench/, which exists so a performance claim can be re-checked by someone who isn't me: alternating A/B arms, medians rather than means, output-equivalence checks before any win is claimed, and a recorded case where the original hypothesis was refuted by the measurements. The changelog is written the same way, including the gates that still do not run.

Interfaces I've designed and shipped. Each link is a live build, not a screenshot.

• Furniture-App-UI — gesture-driven catalog with responsive multi-device viewports. live demo · Flutter, MIT
• employee-ledger — offline-first team directory on Sembast with BLoC. live PWA · Flutter Web
• Coffee-App — OLED-tuned ordering flow, Material 3. live demo · Flutter, MIT
• caretaker — healthcare concierge site, editorial layout. live site · Next.js
• Weather-App — native iOS/macOS weather UI with diurnal theming. · SwiftUI, MIT

These are habits, not slogans — each is visible in the repositories above.

Measure so someone else can re-run it. A benchmark that lives in my /tmp is not evidence. The harness gets committed, with its method and its failure modes.

Separate "faster" from "correct." A fast error response is not a fast search. Output equivalence gets checked before a speedup gets claimed.

Label the maturity. stable, experimental, planned, and known broken are different states, and the README says which one you're looking at.

Caveats go in the document, not the changelog footnote. If a gate doesn't run or a metric is n=5, that belongs where a reader will see it.

Fail closed. When unsure whether something is safe to send, don't send it.

Pin the toolchain. A floating compiler turns a newly published lint into a red build with no code change. Green should mean the same thing on my machine and in CI.

Working fluency, not a claim of mastery in all of it:

Systems    Rust · SQLite / FTS5 · UNIX sockets · POSIX shared memory
Agents     MCP · Pi Agent extensions · T3 Code · tool dispatch
Frontend   TypeScript · React · Next.js · Tailwind CSS
Mobile     Flutter · Dart · Swift · SwiftUI
Tooling    Cargo · Node · GitHub Actions · Xcode · Android Studio

I co-founded Avartana Labs LLP, where we design and ship software products — see the organization on GitHub for what's public there.

Most of my work day is product architecture, delivery, and keeping a small team's engineering quality high. The repositories above are the parts I can show.

_{avartanalabs.com  ·  github.com/Ghost-9}