Inside one IBM research leader’s everyday AI toolkit

Peter Staar does not use one AI assistant. He uses several, often in the same hour.

Staar, Software Manager and Technical Lead of Docling at IBM, works at the intersection of artificial intelligence and large-scale knowledge systems. In practical terms, that means writing and reviewing code, validating ideas against research literature and distilling complex technical material into something usable.

Over the past year, AI tools have become embedded in nearly every step.

“What surprised me most is how naturally these tools have woven themselves into my daily routine,” Staar told IBM Think in an interview. “I didn’t expect to be juggling multiple AI assistants, but each has its strengths for different situations.”

Instead of committing to a single platform, Staar rotates between tools, depending on the task.

For coding, he switches between Claude Code, Codex and IBM’s Bobshell. All three are AI coding assistants that can generate, modify or explain software code. They are typically used from the command line or integrated into development environments to help programmers move faster.

When he wants to validate an idea, test an approach or explore background material, he turns to Claude or ChatGPT, large language models (LLMs) designed to analyze text, answer questions and generate written output. For him, these systems function less as writing tools and more as thinking partners. They help him probe assumptions and quickly surface relevant material.

He has also recently started using Docling, a document-conversion tool, to break down dense technical papers or missed Microsoft Teams meetings into structured summaries. Those summaries can then be fed back into research workflows or used to brief collaborators.

On a typical day, these tools assist with searching for relevant articles and reports, drafting code solutions, formatting reports and preparing slide decks. Tasks that require synthesizing information quickly or producing polished output under time pressure are now routinely handled with AI support.

“It’s become my go-to approach,” Staar said.

The most obvious change, Staar said, is the volume of work he can produce.

“I can now produce three to five times as much output per day as I could before,” he said.

That increase has affected how he plans his work. Projects that once felt ambitious for a week may now be tackled in a day. Tasks that might previously have been postponed are folded into the schedule.

The shift is not only about speed. It has changed what feels feasible within a fixed time frame.

Staar is clear, however, that these gains come with caveats.

For coding, in particular, he believes close oversight of the tools is essential.

“You need to actively supervise the process to prevent unnecessary code bloat and keep the codebase manageable,” he said.

AI-generated code can expand beyond what is needed, introducing extra layers or abstractions that make a system harder to maintain, Staar said. Without supervision, convenience can lead to complexity.

Staar has also noticed a more subtle risk.

“If I don’t stay closely involved, I can lose my connection to the code,” he said. When debugging, that distance can slow him down.

Additionally, the tools still make mistakes, particularly in languages like C++, which require strict syntax and careful memory management. Human review remains essential, Staar said.

For colleagues who are curious but unsure where to begin, Staar’s advice is simple.

“Just start,” he said. “It really isn’t hard to get going, and the productivity gains are remarkable.”