Vespper Launches MCP Tool Letting AI Agents Edit Word Docs via HTML

·1 views

Y Combinator-backed startup Vespper has launched an MCP (Model Context Protocol) server that enables AI agents to edit Word documents more efficiently by presenting them as HTML rather than raw OOXML. Founded by Dudu and Topaz, the team identified the problem while building an AI document editor for pharmaceutical companies, where filling a single clinical study report took around 50 minutes with poor results. Existing approaches — such as using low-level libraries, opinionated MCP tools, or Markdown round-tripping — either waste context tokens on Word mechanics or lose document fidelity. Vespper's solution converts DOCX to HTML as a read-only projection for the agent, then reconciles edits back into the original file, keeping the source document intact. The company claims its tool is currently three times faster, twice as cheap, and more accurate than the nearest alternative.

Read the full story at Hacker News

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

DSPy 2.6 replaces manual prompt engineering with declarative AI pipeline compilation

DSPy 2.6, a framework originally developed at Stanford University, automates the process of prompt creation for large language models by letting developers define task logic rather than writing manual instructions. The framework uses an optimizer called MIPROv2, which takes a small training dataset and an evaluation metric to automatically generate, test, and select the most effective prompts for a given model. This approach addresses a long-standing fragility in handcrafted prompting, where minor model updates or unexpected inputs could cause error rates to spike dramatically and force engineering teams to restart optimization cycles from scratch. DSPy 2.6 structures AI development around modular components and declarative signatures, separating program logic from the underlying prompt text entirely. The release is being positioned as a shift from ad hoc prompt tuning toward a more systematic, compiler-like methodology for building reliable AI pipelines.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Bootstrap Added to Blazor Task Tracker Without Touching App Logic

A developer tutorial series has concluded with its final installment, applying Bootstrap CSS styling to a previously functional but visually plain Blazor Task Tracker app. Bootstrap, a pre-built CSS framework, was used to add layout, spacing, and UI components such as navbars, cards, and buttons without altering any underlying C# logic. The series spanned five parts, beginning with a WPF implementation, progressing through MVVM refactoring, Razor rendering, and Blazor interactivity, before arriving at this styling phase. The final post demonstrates a key software principle: separating presentation from business logic, meaning visual changes required no modifications to the code driving the app's functionality. The complete series serves as a practical walkthrough of transitioning a desktop task management app to a styled, interactive web application using Microsoft's Blazor framework.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

BitNet b1.58 Brings 100B-Parameter AI Inference to Ordinary CPUs

A ternary neural network architecture called BitNet b1.58 reached broad scientific adoption in late September 2026, marking a significant shift in how large AI models are run. The approach encodes neural network weights using only three discrete values — −1, 0, and +1 — requiring roughly 1.58 bits per parameter, which eliminates the need for floating-point matrix multiplication during inference. This means computations reduce to simple integer additions and subtractions, drastically cutting hardware demands. The open-source framework bitnet.cpp has since been stabilized to support mainstream CPUs including Intel Xeon, AMD EPYC, Apple Silicon, AWS Graviton, and edge chips like Snapdragon and RISC-V. Models with up to 100 billion parameters can now run on commercial servers without expensive GPU clusters, potentially democratizing access to large-scale AI inference.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

FlashAttention-4 targets memory bottlenecks in NVIDIA Blackwell B200 GPUs

FlashAttention-4 (FA4), developed by Tri Dao's lab in close collaboration with NVIDIA, has been released as a production specification optimised for the Blackwell B200 and GB200 accelerator architecture. The update addresses critical performance bottlenecks that emerged as large language models such as DeepSeek 4.1 and GPT-6 Astra began operating with context windows exceeding one million tokens. Earlier versions of the algorithm suffered synchronisation stalls and memory bandwidth saturation on Hopper-generation hardware, wasting up to 32% of available clock cycles. FA4 resolves this by introducing Asymmetric Kernel Pipelining via a hardware Tensor Memory Accelerator, splitting workloads between dedicated Producer and Consumer Warps running asynchronously. The release also adds native support for 4-bit quantisation formats FP4 and MXFP4, enabling fuller utilisation of the B200's 20 PFLOPS FP4 compute capacity and 8 TB/s memory bandwidth.

0 comments Read more at DEV Community