Clelia Astra Bertelli

My latest setup for agentic engineering in personal projects includes Copilot and Pi powered by Kimi K2.6.

Of course, I want to track costs, usage and input/outputs, so I adapted fuseproxy, the Langfuse-instrumented proxy I built for th Anthropic Messages API, to natively integrate with Pi.

Coding agents might feel like magic, but they aren’t 🪄 That’s exactly why I’ve been eager lately to understand how they work under the hood, using the most natural approach I could think of: observability.

I also built instrument-claude, an observability tool that turns Claude Code's hook events into OTel-compatible traces, spans, and events, providing an overview of everything happening within a conversation turn or across an entire session.

Coding agents might feel like magic, but they aren’t 🪄 That’s exactly why I’ve been eager lately to understand how they work under the hood, using the most natural approach I could think of: observability.

Coding agents shouldn’t be black boxes ❌ If you’re calling models through APIs and can configure provider settings (base URL, API key, etc.), I highly recommend putting a local instrumented API proxy in front of them. I built one for the Anthropic Messages API:

Best part? Everything runs locally and fully self-hosted with Docker Compose. This setup gives me: - visibility into what the agent actually sends and receives - insight into tool calls and tool outputs - token + cost tracking - latency monitoring

Coding agents shouldn’t be black boxes ❌ If you’re calling models through APIs and can configure provider settings (base URL, API key, etc.), I highly recommend putting a local instrumented API proxy in front of them. I built one for the Anthropic Messages API:

Do you actually own your document parsing infrastructure? 👀 At LlamaIndex, we wanted to make that easier, so we built 𝗹𝗶𝘁𝗲𝗽𝗮𝗿𝘀𝗲-𝘀𝗲𝗿𝘃𝗲𝗿, a lightweight HTTP backend built on top of LiteParse that can parse and generate page screenshots from PDFs, images, and Office documents🦙

We also included ready-to-use examples for rate limiting, caching, and OpenTelemetry-compatible traces and metrics collection with tools like Redis, Jaeger, Prometheus and Grafana🔭

Do you actually own your document parsing infrastructure? 👀 At LlamaIndex, we wanted to make that easier, so we built 𝗹𝗶𝘁𝗲𝗽𝗮𝗿𝘀𝗲-𝘀𝗲𝗿𝘃𝗲𝗿, a lightweight HTTP backend built on top of LiteParse that can parse and generate page screenshots from PDFs, images, and Office documents🦙

I have one big problem with agentic engineering: I want agents to operate autonomously, but I also want granular, reversible control over every change they make. I could solve this by committing every intermediate step to Git, but that would completely pollute my repo history.

With aggit, my agents can stash intermediate work, create branches safely, restore previous states, and back up artifacts to my RustFS server with multi-instance replication support, all without touching the main Git repository.

I have one big problem with agentic engineering: I want agents to operate autonomously, but I also want granular, reversible control over every change they make. I could solve this by committing every intermediate step to Git, but that would completely pollute my repo history.

For 2026, I decided I'm going to build all of my side projects with Rust when possible🦀 So here's the first one: 𝗳𝗶𝗹𝗲-𝘀𝘁𝗼𝗿𝗮𝗴𝗲, a self-hostable file storage application built on a modern open source stack, with a full Rust backend and a @react.dev + @vite.dev frontend🚀 With file-storage, you get:

🌐 Axum REST APIs to interact with Redis and Qdrant ➕ Observability with Jager and @opentelemetry.io, rate limiting with Memcached 👩‍💻 And, as always, the code is open source if you want to customize it: github.com/AstraBert/f...

For 2026, I decided I'm going to build all of my side projects with Rust when possible🦀 So here's the first one: 𝗳𝗶𝗹𝗲-𝘀𝘁𝗼𝗿𝗮𝗴𝗲, a self-hostable file storage application built on a modern open source stack, with a full Rust backend and a @react.dev + @vite.dev frontend🚀 With file-storage, you get:

Scaling document pipelines is harder than it looks, and not because of tooling, but mostly because of orchestration⚙️ That's why I teamed up with @render.com and focused on building a system where each component does one job well, and scales independently:

The interesting part here is not the single step, but how we orchestrate all the moving parts without introducing bottlenecks. Read the full breakdown 👉 render.com/blog/buildi...

Scaling document pipelines is harder than it looks, and not because of tooling, but mostly because of orchestration⚙️ That's why I teamed up with @render.com and focused on building a system where each component does one job well, and scales independently:

The LlamaParse MCP got a new face, and it is now easier than ever to run document processing workflows from your agents🚀

We refactored our MCP to have: - Direct integration with our Parse, Classify and Split services🦙 - A smoother authentication flow using WorkOS🔒 - Seamless support for file uploads⬆️ - Observability, rate-limiting and fast deployments with @vercel.com and @axiom.co

I'll be speaking at the Applied AI Conf organized by Tech Europe!🇪🇺 On the 28th of May, I'll be in Berlin to give my talk about how you can build reliable document agents by leveraging harness engineering with the building blocks provided by @llamaindex.bsy.social 🦙

Thanks to everyone who helped in getting me on board with this! PS: link for tickets in the next comment👇

I've been building sunbears, a typescript library for CSV data loading written in Rust🦀

📝I decided to write a blog post about it, mostly to document the building journey for myself, but also to share my thinking and coding process around sunbears: clelia.dev/blog/2026-0... Enjoy!✨

ParseBench is here!📊 We’ve just released ParseBench, an open benchmark + dataset for evaluating document parsing at scale.

What makes it different? ParseBench optimizes for semantic correctness, not exact text matching. That means evaluating whether parsed outputs are actually useful for humans and AI agents making downstream decisions 🔍

ParseBench is here!📊 We’ve just released ParseBench, an open benchmark + dataset for evaluating document parsing at scale.

I'm building 𝘀𝘂𝗻𝗯𝗲𝗮𝗿𝘀, a CSV data loader library for TS written in Rust🦀 In Node, it can read a file with 1.000.000 rows in 0.3s, write the same amount of rows in 0.15s, respectively 4x and 2x faster than the `csv` package⚡

sunbears uses 𝘋𝘢𝘵𝘢𝘍𝘳𝘢𝘮𝘦 as its primary data structure, a columnar format with strict typing, null and NaN filtering, and convenient column-to-array transformations📄 📦 Get started now: 𝘯𝘱𝘮 𝘪𝘯𝘴𝘵𝘢𝘭𝘭 @𝘤𝘭𝘦-𝘥𝘰𝘦𝘴-𝘵𝘩𝘪𝘯𝘨𝘴/𝘴𝘶𝘯𝘣𝘦𝘢𝘳𝘴