
How I built an LLM eval harness that took AI output from 24% to 97% valid
A blind pairwise judge, a pinned prompt registry and one harness bug that silently inverted our model rankings. A practical guide to LLM evals for code generation.
Full-stack developer shipping LLM pipelines, AI evals and MCP servers on robust backends, clean frontends and smooth deployments for global clients.
Adaptable and quick-learning Full Stack & AI Developer with over 5 years of experience. Proficient in designing, developing, and deploying AI-powered products, web applications and backend services, particularly for international clients.
Based in Nepal, I've had the privilege of working with clients from around the globe, delivering scalable solutions that make a real impact. My expertise spans from architecting robust backend systems to crafting intuitive user interfaces.
Lately my focus is AI engineering: multi-model LLM pipelines with Claude and OpenAI, eval harnesses that prove a model is ready to ship, sandboxes that keep AI-generated code safe, and MCP servers that connect products to AI assistants. I build with Claude Code in spec-driven, agentic workflows.
I thrive on challenges and continuously seek to enhance my skills with emerging technologies. Whether it's leading development teams or diving deep into complex technical problems, I bring dedication and innovation to every project.
A comprehensive toolkit built through years of hands-on experience and continuous learning, now centered on shipping AI.
A journey through innovative companies and challenging projects that shaped my expertise.
From AI products to marketplaces and games, built for clients across the globe.
What I'm learning while shipping AI products.

A blind pairwise judge, a pinned prompt registry and one harness bug that silently inverted our model rankings. A practical guide to LLM evals for code generation.

When a model writes the React your users see, treat its output as hostile. Here's the defence in depth we use: a shared contract, AST validators, a quality gate and a cross-origin sandbox.

What it took to let customers connect Claude and ChatGPT to their own data safely: OAuth 2.1 with PKCE, dynamic client registration, scoped tools, live permissions and rate limits per connection.
Ready to bring your next project to life? Let's discuss how we can work together.