AIware Bootcamp - Europe 2025

Opening of the Bootcamp

Organizers: Gopi Krishnan Rajbahadur, Ahmed E. Hassan, Maliheh Izadi

From Copilots to Agentic SE Teammates: Engineering the SE 3.0 Era with AI-Native Workflows

Ahmed E. Hassan

Session Outline:

• What is SE 3.0? – A new paradigm of Agentic Software Engineering where autonomous, goal-oriented AI collaborators evolve from passive copilots to proactive teammates.
• Why SE 2.0 Isn’t Enough – Limitations of current FM-powered copilots: developer overload, fragmented workflows, and lack of agentic behavior and trust.
• The SE 3.0 Agentic Stack:
• Teammate.next – Adaptive, personalized agentic partners capable of shared goals and proactive decision-making.
• IDE.next – Intent-first, conversation-driven development environments designed for agent collaboration.
• Compiler.next – Multi-objective code generation powered by agents optimizing for quality, performance, and compliance.
• Runtime.next – SLA-aware, distributed execution environments orchestrating both human and agent workflows.
• Introducing AIDev – The first large-scale dataset capturing AI-agent-generated pull requests across 61K+ repositories.
• What AIDev Reveals – Agentic contributions at scale and speed—yet facing acceptance, coordination, and complexity gaps.
• Beyond Benchmarks – Why real-world agentic data (AIDev) exposes gaps that synthetic benchmarks like SWE-bench miss.
• Research & Engineering Opportunities – Benchmarking agent trust, modeling collaboration, and embedding governance into Agentic SE workflows.
• Call to Action – Building metrics, tooling, and safeguards to fully realize Agentic SE 3.0 in practice.

Building High-Quality & Trustworthy FMware

Dayi Lin

Session Outline:

• Overview of AIware
• Introduction to the trustworthiness of software and AI
• How do you select the right model?
• How to benchmark a model?
• How to write a good prompt?
• How to prevent hallucination?
• How to debug prompts?
• How to prevent getting or causing harm – Guardrails
• How to ensure compliance in dataflow?
• How to conduct quality evaluation?
• How to interact with the users?
• How to operationalize the application?

Morning Break

Practical Evaluation of LLMs for Code

Maliheh Izadi

Session Outline:

• Practical application of open-weight large language models for code generation
• Evaluation of model performance in live software development environments
• Analysis of failed cases and insights from these failures
• Overview of AI4SE benchmarking approaches
• Discussion of existing challenges in AI for software engineering

Beyond Vibe Coding: Leveraging Foundation Models to Infer Intent

Keheliya Gallaba

Session Outline:

• Why Requirements Engineering is the Bottleneck – Aligning software with true user needs starts with understanding intent.
• From Vibe Coding to Real Engineering – Critiquing the throwaway culture of prompting and patching.
• What Are Intents (Really)? – Distinguishing goals, requirements, and the hidden rationale that underpins them.
• Inferring Intents with Foundation Models – Using Theory of Mind (ToM), iterative refinement, and dialog context to bridge ambiguity.
• Conversational Interfaces for Requirement Elicitation – Proactively clarify requirements, resolve contradictions, and avoid premature solutions.
• Evaluating Intent Quality – Rubrics for accuracy, relevance, completeness, realizability, and collaboration support.
• Real-World Gains Over Baselines – Evidence that FM-inferred requirements deliver better grounding and satisfaction.
• Known Challenges – Tackling sycophancy, hallucination, and compute cost.
• What’s Next for Requirements Engineering – Reasoning-augmented models, multi-modal inputs, and non-chat interfaces.
• Call to Action – Foundation Models shouldn’t just write code—they must understand why the code exists.

Lunch Break

Lunch not provided: participants are encouraged to explore the local cuisine.

From Chaos to Compliance: Engineering Legal Assurance in the FMware Era

Gopi Krishnan Rajbahadur

Session Outline:

• The Legal Risk Landscape for FMware – Unpack rising lawsuits in AI-powered software.
• Why Traditional Compliance Doesn’t Scale – Gaps in software engineering, OSPO practices, and legal frameworks.
• The Dataset License Dilemma – Ambiguity in public dataset licenses poses real risks.
• Introducing LicenseGPT – A fine-tuned FM on 500 expert-annotated licenses for automated compliance.
• Human-in-the-Loop Legal AI – IP lawyers using LicenseGPT as an accelerator, not a replacement.
• From Licensing to Traceability – Emergence of AIBOM and FMwareBOM for provenance tracking.
• Compliance as Engineering Practice – Embedding legal assurance into SE pipelines.
• The Role of Standards and Open Source – OSPOs, SPDX, and the open source community leading compliance innovation.
• Call to Action – Treat legal risk in FMware as a core engineering discipline.

Visual Quality Assurance Using Vision-Language Models

Cor-Paul Bezemer

Session Outline:

• Why Visual QA Matters – Ensuring user trust and correctness in dynamic, visual interfaces.
• Limits of Manual and Snapshot Testing – Current methods are brittle and labor-intensive.
• Non-Determinism in Modern Software – AI-driven behavior makes pixel-perfect matching unreliable.
• When Bugs Aren’t Bugs – Why visual differences don’t always indicate regressions.
• The Commonsense Gap – Many visual issues require human-like reasoning.
• Vision-Language Models for Visual QA – Using multimodal models to understand and evaluate visual intent.
• Detecting Subtle Visual Failures – Identifying layout, style, and semantic inconsistencies.
• Natural Language Explanations – Producing interpretable reports for visual issues.
• Challenges and Open Questions – Hallucinations, evaluation criteria, and integration hurdles.
• Toward Smarter Testing Pipelines – Embedding VLMs into scalable, automated QA workflows.

Afternoon Break

From Issues to Commits: Autonomous Agents for Smarter Software Traceability

Mehdi Keshani

Session Outline:

• What is issue-to-commit link recovery, and why does it matter?
• What makes it challenging in real-world projects?
• Where do existing AI/LLM methods fall short?
• How does an agentic approach solve these challenges?
• How well does it perform against state-of-the-art?
• Key lessons from agentic issue-to-commit link recovery.

Panel: Agentic SE & Open Source in the FMware Era – Toward Trustworthy, Context-Aware FMware

Cor-Paul Bezemer, Day Lini, Keheliya Gallaba, Maliheh Izadi, Gopi Krishnan Rajbahadur
Moderator: Mehdi Keshani

Group Picture & Closing of the Bootcamp

Organizers: Gopi Krishnan Rajbahadur, Ahmed E. Hassan, Maliheh Izadi