AI Architect
I design enterprise AI deployments at ThoughtSpot and build the agents and tools that make customer delivery faster.
LinkedIn GitHub (personal) GitHub (work) bibek.shrestha.work@gmail.com
Before you install an MCP server, you want to know if it is secure and still maintained.
I built an open catalog that indexes MCP servers and scores each one on security, maintenance, efficiency, documentation and compatibility, with real CVE scanning and tool-poisoning detection. It also ships as an MCP server on npm (7 tools, stdio and Streamable HTTP), so you can search and compare servers from inside an AI client.
50,000+ servers scored. Listed on Smithery.
I wanted to share photos at full resolution, with no account needed to view them.
I designed, built and launched it solo in about 226 commits, a photo portfolio and private sharing platform where images are never used for AI training. Next.js 16 on Cloudflare Workers, Neon Postgres with Drizzle, Better-Auth, and presigned direct uploads to R2 with BlurHash placeholders and EXIF extraction.
Live in production.
Plan a trip by talking about it, then get a day-by-day itinerary on a map.
Claude gathers trip details in chat, then Gemini 2.5 Flash with Maps grounding generates an itinerary that is geocoded and route-optimized. Past trips feed back in through RAG: gemini-embedding-001 vectors in pgvector, retrieved by cosine similarity into both prompts. Every AI call goes through Supabase Edge Functions so keys stay server-side, and every LLM JSON response is validated with Zod.
I also built an eval framework for it: 55 scenarios across 30 categories (161 graded checks), graders calibrated against a hand-graded gold set, adversarial-user tests, and a baseline-diff gate that blocks prompt regressions.
5,982 tests passing, 0 type errors. About 1,220 commits and 268 merged PRs. In private beta.
A Claude Code skill that builds ThoughtSpot custom charts end to end. It renders each chart in a real browser through Playwright and critiques its own output until the render matches, so customers can build their own charts.
Building a custom chart went from days, if not weeks, to minutes.
An open-source Claude Code plugin that generates a CEO-led agent org for any repo: tailored agents, an orchestrator, memory, guard hooks, evals, and an autonomy level the org earns on scorecard evidence. Only builder agents can edit code, and every miss a human catches becomes a test the agents must pass.
Adopted in Photobari: 29 merged agent PRs, 0 reverts.
A community web control plane for ThoughtSpot administrators, with a Python/FastAPI backend and a Next.js frontend. Bulk archive, delete, share and ownership-transfer jobs run with dry-run previews and a TML backup first, and it installs with one line.
Lead contributor, 68 of 92 commits, shaped by feedback from field users.
A zero-build sandbox for the ThoughtSpot Visual Embed SDK, with a hardened Express service that mints trusted-auth tokens server-side. A multi-agent engineering org maintains it, and a PR merges only when CI is green and review finds no confirmed bug.
86 commits. Security-critical changes are held for human approval.
An open-source JavaScript chart library: 28 chart types from one JSON spec, about 87 KB gzipped, zero runtime dependencies, accessible and server-rendered by default, and able to handle up to 1M rows.
Published on npm as mayacharts.
AI Architect, Customer Outcomes 2026 to now
Solutions Architect 2025 to 2026
Lead Solutions Engineer 2023 to 2025
Solutions Engineer 2021 to 2023
I build with Claude Code, and I run it with the discipline I would expect from any engineering team: a spec first, automated tests, CI gates, adversarial review, and human approval on anything security-critical.
My background is seven years of customer-facing solutions engineering and architecture, and I have not held a software engineer title. So I let the projects above carry the proof, and most of them are public or live so you can check them yourself.