Uploaded June 2026 | Updated September 2026, 2 weeks ago
Marcin Dubel, Appsilon's Head of Technology, sat down with AI Architect Jakub Wąsala to talk about what it really takes to make pharma data usable for AI. Not the demo version where you point a model at a folder and hope for the best, but the work that happens before any of it pays off.
They get into why so many AI projects stall early (the data foundation usually isn't there), how vector databases and semantic search make a pile of documents searchable in seconds, and why handing an agent access to all your files tends to backfire. They also cover why evaluation and observability need to be part of the system from day one, not something added after users start complaining things broke.
One bit worth stealing: before writing any code, ask people what questions they actually want the system to answer. That usually tells you what work has to come first.
If you work in pharma data, platform engineering, or you're the person everyone keeps asking to "add some AI," there's something here for you.
Mentioned in the conversation:
Validated AI for Pharma Summit:
appsilon.com/validated-ai-for-pharma-summit
Advanced Patterns for Cost- and Time-Efficient Development with Copilot, Claude Code, or Codex:
appsilon.com/post/advanced-patterns-for-cost-and-time-efficient-development-with-copilot-claude-code-or-codex
More about Appsilon:
appsilon.com
#PharmaAI #ClinicalData #RAG #DataEngineering
Marcin Dubel, Appsilon's Head of Technology, sat down with AI Architect Jakub Wąsala to talk about what it really takes to make pharma data usable for AI. Not the demo version where you point a model at a folder and hope for the best, but the work that happens before any of it pays off.
They get into why so many AI projects stall early (the data foundation usually isn't there), how vector databases and semantic search make a pile of documents searchable in seconds, and why handing an agent access to all your files tends to backfire. They also cover why evaluation and observability need to be part of the system from day one, not something added after users start complaining things broke.
One bit worth stealing: before writing any code, ask people what questions they actually want the system to answer. That usually tells you what work has to come first.
If you work in pharma data, platform engineering, or you're the person everyone keeps asking to "add some AI," there's something here for you.
Mentioned in the conversation:
Validated AI for Pharma Summit:
appsilon.com/validated-ai-for-pharma-summit
Advanced Patterns for Cost- and Time-Efficient Development with Copilot, Claude Code, or Codex:
appsilon.com/post/advanced-patterns-for-cost-and-time-efficient-development-with-copilot-claude-code-or-codex
More about Appsilon:
appsilon.com
#PharmaAI #ClinicalData #RAG #DataEngineering
