Systems Architect · AI Integration Engineer

B.Tech CSE (VIT Vellore) · Kathmandu, Nepal · Nova LinkedIn · LeetCode · burn live docs · limen live docs

I build local-first AI tools, and I put the fixes back into the projects I actually use instead of keeping them in my own fork.

• 83 pull requests merged into other people's projects in September 2026.
• 64 of them into Laya, a 28,000-star AI decision engine. I'm its biggest contributor outside the owner: 117 contributions to its graph, against 37 for the next name.
• Around 20,000 lines added, nearly all of it in that one project.
• 64 more pull requests open across 32 repositories, waiting on review.

Every number here was pulled from the public GitHub API on 29 Sep 2026. The queries are at the bottom — re-run them, and if a figure has drifted, the query is still the truth.

Laya — 64 merged. Laya answers with a single forward pass instead of generating token by token, so speed and correctness are the product. I made it batch across every backend instead of handling one request at a time, gave its TypeScript SDK the same features as the Python core, built the ways people actually reach it (command line, MCP server for agents, LangChain), and made its benchmarks report what they really measured — which commit a score came from, and whether the model ran on the GPU or quietly fell back to the CPU. A few examples: batching, the CLI, honest health checks.

Documentation — 15 merged. 6 into freeCodeCamp curriculum, 5 into public-apis (dead links removed, broken tables repaired), 4 into MDN.

Smaller fixes — 4 merged. A savings counter that printed $0.00 instead of pricing itself in headroom, a float-rounding bug in archify's layout solver, a false positive in flake8-bugbear, and a missing test for nested fixture discovery in ProtocolCanary.

64 open pull requests. Where they are: Laya 8 · pear-desktop 5 · paperclip 5 · odysseus 5 · mempalace 4 · the remaining 37 across 27 repos

17 repositories, 75 stars between them. The six that matter:

• Aniflow ★9 — turns webtoons into narrated videos: object detection, multi-voice narration, automated editing.
• burn ★8 — tells you what your coding agents actually cost. Reads the logs Claude Code, Codex, OpenCode and Gemini CLI already leave on disk. No account, no network. Live docs
• limen ★7 — shows how much of your prompt is boilerplate your agent re-sends every call, and cuts it. Live docs
• stratmate ★7 — planning and timing math for mobile game alliances.
• agentvault ★6 — a permission firewall for AI agents. On its own benchmark: 103 of 103 attacks blocked, 8 of 8 normal operations allowed.
• ulp ★6 — an open file format for agent usage and cost, mergeable between machines with no server.

23 pull requests were closed without ever being merged, and that belongs in the record too. Ten of them were duplicates I opened against the same files in tldr-pages — the maintainers closed the whole set. The rest split across MDN (2), freeCodeCamp (2), Laya (2), JabRef (2) and five other repos.

# merged PRs (84 total, 83 excluding one to my own stratmate)
gh api "search/issues?q=is:pr+author:aashish254+is:merged&per_page=100"

# open PRs (64)
gh api "search/issues?q=is:pr+author:aashish254+is:open&per_page=100"

# closed without merging (23)
gh api "search/issues?q=is:pr+author:aashish254+is:closed+is:unmerged&per_page=100"

# lines landed: additions/deletions/changed_files for each merged PR
gh api repos/<owner>/<repo>/pulls/<number> --jq '{additions,deletions,changed_files}'

# contributions calendar (509 over 53 weeks)
gh api graphql -f query='{viewer{contributionsCollection(
  from:"2025-09-29T00:00:00Z",to:"2026-09-29T23:59:59Z"){
  contributionCalendar{totalContributions weeks{contributionDays{contributionCount date}}}}}}'

# my repositories and their stars
gh api users/aashish254/repos --paginate --jq '.[] | select(.fork==false)'

Engineering notes: system architecture & API orchestration patterns · developer onboarding & contribution practice