Reimagining AI: Building applications and data on AWS

 Decades of technical debt, aging applications and siloed data are slowing modernization and starving AI of the trusted information it depends on. Gartner® predicts that “through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data.”*

As the widening gap between AI ambition and operational reality becomes a competitive risk, the answer is a connected foundation that modernizes the digital estate, industrializes trusted data and AI, and scales agentic operations with confidence.

Agentic AI is where the work is heading. Amazon Bedrock AgentCore is a platform for building and operating production AI agents across frameworks and models and securing them at scale.

You can build a useful agent or prototype relatively fast. What slows you down is connecting it to enterprise systems, giving it controlled access to tools and data, evaluating behavior, operating it reliably and improving it over time. AgentCore handles the heavy lifting that slows you down. Our role is to help you use this capability at scale with the governed operating model, engineering discipline and delivery capacity needed for complex enterprises.

A practical foundation for AI starts with the systems and data you already have. We help you modernize applications, reduce technical debt, connect fragmented data and improve software development with AI.

In our experience, AI-augmented development drives up to 50% productivity gains in application modernization, and our industrialized modernization factory approach delivers up to 40% faster time to market.

The architectural shift is also important. Moving from monolithic systems to API-first, cloud-native applications accelerates release cycles by 25% to 50% and improves application responsiveness by up to 40%.

These gains are the prerequisites for AI that can reason within live, connected systems.

AI and Mainframe Modernization: A Force Multiplier, Not a Replacement for Expertise

 Generative AI has changed the economics of mainframe modernization. Much of the legacy estate is built from well-understood, repetitive constructs: COBOL programs, IMS databases and transactions, JCL job streams. Because these follow consistent conventions, AI tools can analyze them, document them, and translate them into modern languages and platforms far faster than a team of engineers working by hand. Work that once took months of tedious, line-by-line conversion can now be compressed dramatically.

But speed at the translation layer is not the same as solving the modernization problem. Code conversion has always been the most visible part of these programs, not the hardest. The real difficulty lies in what the code means to the business, and how it connects to everything around it. AI can convert syntax reliably; it cannot tell you why a particular rounding rule exists, which regulatory requirement a batch job quietly satisfies, or what downstream system breaks if a nightly file arrives in a slightly different format.

What a Hybrid Approach Looks Like

The programs that succeed tend to divide the work along natural lines:

• AI handles technical translation. Pattern-based conversion of standard components, code analysis, dependency mapping, test generation, and first-draft documentation are all strong fits for automation.
• Domain experts own business logic. People who understand the underlying processes, whether claims adjudication, interest calculation, or settlement, must verify that converted behavior matches real-world intent, not just that the code compiles and runs.
• Architecture favors incremental replacement. Rather than betting everything on one big-bang cutover, successful teams migrate in phases, running old and new systems in parallel, validating outputs against each other, and retiring legacy components piece by piece as confidence grows.
  
  

It's worth being explicit about what AI does not change. Decades-old mainframe environments accumulate custom utilities, one-off scripts, and business rules embedded in code that no one documented and whose authors retired long ago. These were the hidden obstacles that derailed earlier modernization attempts, and they remain just as hidden today. Uncovering them still requires reverse-engineering, careful validation against production behavior, and often manual refactoring.

The practical takeaway: budget time, talent, and money for this discovery and validation work from the start. AI will shorten the timeline for the routine portions, but the difficult, undocumented parts of the system will demand the same rigor they always have.

AI Development Life Cycle is different from Vibe Coding

 

AI Development Life Cycle (AI DLC)

A typical AI DLC includes these stages:

1. Problem definition – Define the business problem and success metrics.
2. Data collection – Gather and prepare training and evaluation data.
3. Data preprocessing – Clean, transform, and label the data.
4. Model selection and training – Choose algorithms or foundation models and train or fine-tune them.
5. Evaluation – Measure accuracy, robustness, fairness, latency, and cost.
6. Deployment – Release the model into production.
7. Monitoring – Track performance, drift, failures, and user feedback.
8. Maintenance – Retrain, update, and improve the system over time.
   
   

   Vibe coding

   Vibe coding is a development workflow rather than a lifecycle. It typically looks like:

   1. Describe what you want in natural language.
   2. AI generates code.
   3. Test the result.
   4. Ask the AI to fix or improve it.
   5. Repeat until you're satisfied.

   The AI does much of the implementation, while the developer focuses on defining requirements, reviewing outputs, and making decisions.
   
   
   

   They're complementary rather than competing:

   • Use vibe coding when you want to prototype or build software quickly.
   • Use the AI Development Life Cycle when you're creating an AI product that needs to be reliable, secure, maintainable, and suitable for production.

AI-Native Software Development Needs More Than AI Coding

 The conversation around AI in software engineering often focuses on how quickly code can now be generated. AI assistants can write functions, create tests, generate documentation, and even fix bugs in a fraction of the time it once took developers. But speed alone doesn't define success.

As organizations embrace AI-native software development, a more important question is emerging: How do we ensure AI creates measurable business value?

The answer lies in treating AI as an economic resource rather than simply another development tool.

In an AI-first Software Development Life Cycle (SDLC), developers spend less time writing code and more time defining requirements, validating outcomes, and governing AI-generated work. AI becomes the execution engine, while humans focus on strategy, architecture, and decision-making.

This shift also demands a new way of measuring productivity. Traditional metrics such as developer hours, story points, or lines of code become less meaningful when AI can complete technical tasks in minutes. Instead, organizations need to evaluate outcomes—delivery speed, software quality, AI utilization, operating costs, and business impact.

Another emerging challenge is orchestration. With multiple AI models and agents available, selecting the right model for the right task becomes critical. Efficient AI usage isn't just about capability—it's about balancing performance, quality, and cost.

The organizations that succeed with AI won't simply be those using the most AI tools. They'll be the ones that establish governance, optimize AI investments, and continuously measure the value AI delivers across the development lifecycle.

The future of software engineering isn't just AI-powered—it's AI-managed. As AI becomes a core part of software delivery, success will increasingly depend on combining intelligent automation with sound economic decision-making.

Cloud modernization: AI is exposing the limits of legacy systems

 

For years, modernization was largely viewed through the lens of efficiency.

 

Reduce technical debt. Improve performance. Lower costs. Those outcomes still matter. But in the era of AI, they're no longer the primary reason organizations modernize.

What we're seeing across industries is a fundamental shift in expectations. Leaders want AI to move beyond pilots and proofs of concept. They want it embedded into business processes, customer experiences, and decision-making at scale. That requires more than access to models. It requires modern applications, modern data architectures, and cloud-native foundations capable of operationalizing AI across the enterprise. Yet many organizations remain constrained by legacy environments that were never designed for this reality.

 

The result? AI initiatives that demonstrate promise but struggle to scale. The organizations creating the most value from AI are not simply layering AI onto existing environments. They are reimagining the applications, workflows and data foundations that enable AI to deliver meaningful business outcomes. Modernization is no longer a technology initiative. It's increasingly becoming the pathway to AI value.

 

In my latest blog, I explore why application modernization has become one of the most important priorities for organizations looking to move from AI experimentation to enterprise impact.

https://tinyurl.com/ysxn5njn 

Spec-Driven Development for the SDLC

 As software systems become increasingly complex and AI-assisted development becomes mainstream, organizations are rethinking how they manage the Software Development Life Cycle (SDLC). One emerging approach gaining momentum is Spec-Driven Development (SDD), where specifications become the primary source of truth for the entire development process.

Traditionally, software projects rely on multiple artifacts such as business requirements documents, user stories, design diagrams, code repositories, test cases, and operational runbooks. Over time, these artifacts often drift out of sync, creating ambiguity, rework, and communication gaps between business stakeholders, architects, developers, testers, and operations teams. The result is a fragmented SDLC where teams spend significant effort interpreting intent rather than delivering value.

Spec-Driven Development addresses this challenge by placing machine-readable specifications at the center of the lifecycle. These specifications describe not only business requirements but also functional behavior, non-functional requirements, interfaces, acceptance criteria, security policies, and architectural constraints. Rather than serving as static documentation, specifications become living artifacts that drive downstream activities.

When specifications are treated as the source of truth, every stage of the SDLC can be aligned to the same authoritative definition. Architects use specifications to generate solution designs. Developers use them to create implementation plans and generate code. Test teams derive automated test cases directly from acceptance criteria. Operations teams leverage the same specifications to define deployment and monitoring requirements. This creates a continuous thread of traceability from business intent to production deployment.

The rise of AI-powered development tools further amplifies the value of SDD. Large Language Models and software agents perform best when provided with clear, structured, and unambiguous instructions. Machine-readable specifications enable AI agents to generate code, create tests, validate compliance, and even suggest architectural improvements while maintaining alignment with original requirements. Frameworks such as GitHub Spec Kit, Agent OS, and other specification-centric approaches are emerging to support this model.

Beyond productivity gains, SDD improves governance and quality. Changes made to specifications automatically propagate through the development workflow, reducing inconsistencies and enabling impact analysis. Teams gain greater confidence that delivered software reflects approved requirements, while auditors and stakeholders benefit from enhanced transparency and traceability.

As organizations move toward AI-native software engineering, specifications are evolving from supporting documents into executable knowledge assets. In this new paradigm, Spec-Driven Development transforms specifications into the single source of truth, creating a more consistent, automated, and intelligent SDLC that connects strategy, design, development, testing, and operations.

The role of the Forward Deployed Engineer

 A forward deployed engineer (FDE) represents a new breed of technologist who operates at the intersection of software engineering and real-world business execution. Unlike traditional engineers who build products from within the confines of internal teams, FDEs work directly with customers, often embedding themselves within client environments to ensure that technology solutions function effectively in real-world conditions. (https://futurense.com/blog/fde-forward-deployed-engineers)

The core responsibility of a forward deployed engineer is to bridge the gap between what a product is designed to do and how it performs in practice. They customise, integrate, and optimise complex systems—particularly in areas such as cloud computing, enterprise platforms, and artificial intelligence—to align with specific customer needs.  Their work typically spans the entire lifecycle of a solution, from initial requirements analysis and design to deployment, troubleshooting, and ongoing optimisation. 

What makes the role distinctive is its hybrid nature. FDEs must combine deep technical expertise with strong communication skills, as they collaborate closely with both engineering teams and business stakeholders. They operate in dynamic, often ambiguous environments, solving problems in real time and adapting solutions as requirements evolve. 

In today’s era of rapidly evolving AI and enterprise systems, forward deployed engineers are increasingly critical. They ensure that innovation translates into tangible business outcomes, making them essential to successful digital transformation initiatives.

Make it in Germany

 Thinking about starting your professional life abroad? Whether you bring professional expertise and commitment, wish to pursue vocational training, or plan to study here – Germany is looking for people just like you.

"Make it in Germany" is the German Federal Government's official information portal for international skilled workers. 

https://www.make-it-in-germany.com/en/ 

Whether you are abroad or already in Germany, find out here how to encounter job opportunities in Germany.

https://www.make-it-in-germany.com/en/working-in-germany/job/looking-for-job

Token Efficiency: The Next Frontier in AI Architecture

 As organizations move from AI experimentation to enterprise-scale deployment, attention is shifting from model capability alone to the economics of operating AI at scale. One of the most important factors in this equation is token efficiency—the ability to deliver the desired business outcome while minimizing unnecessary interactions with large language models.

Many AI solutions incur avoidable costs through oversized prompts, redundant processing, excessive context sharing, or the use of highly capable models for relatively simple tasks. While these inefficiencies may seem minor in isolation, they can significantly impact performance, response times, and operating costs when multiplied across thousands or millions of requests.

Forward-thinking architectures address this challenge by intelligently managing context, reusing previously generated insights, matching workloads to the right models, and providing only the information needed to complete a task. The goal is not simply to reduce token usage, but to optimize the balance between cost, speed, and quality.

As AI becomes embedded in core business processes, token efficiency is evolving from a technical consideration into a strategic architectural principle. Organizations that build with efficiency in mind will be better equipped to scale AI adoption sustainably while maximizing return on investment.

AI’s Real Bottleneck Isn’t the Model—It’s the Architecture

 AI is no longer constrained by model capability—it’s constrained by the environment in which it operates. As AI systems mature, the real challenge has shifted to control: over compute, data access, security, and where workloads physically run. Traditional cloud architectures, built for centralized and borderless data flows, are increasingly misaligned with these needs.

At the core of this shift is data jurisdiction. While data can technically move, it cannot move freely in the ways AI demands. Continuous data access and fluid movement are fundamental to AI performance, yet regulatory, sovereignty, and locality constraints are now dictating where data resides, where models execute, and how systems are governed. Architecture is no longer just technical—it is geopolitical.

Most organizations recognize this shift, but few are acting decisively. While over 95% acknowledge the importance of private and sovereign AI, only about one-third are making near-term investments. This gap is creating a widening divide.

Leaders are moving early—redesigning infrastructure, governance, and operating models to accommodate these constraints. As a result, they are scaling faster, moving beyond pilots while others remain stuck in experimentation.

Ironically, pursuing “sovereignty” doesn’t reduce dependency—it increases it. Private and sovereign AI depend on tightly coordinated ecosystems across partners, platforms, and layers. Integration complexity is now the biggest blocker, cited by over half of organizations.

The takeaway: AI advantage will not come from better models alone, but from better-designed, jurisdiction-aware systems.

Token Efficiency is Becoming the New Enterprise AI Advantage

 

Most conversations around artificial intelligence focus on model capabilities—larger models, better reasoning, and more sophisticated outputs. However, as AI adoption scales across enterprises, a more fundamental constraint is emerging: efficiency. Specifically, how effectively organizations manage tokens—the basic units of input and output in large language models—has become a critical determinant of success.

Tokens are not just a technical construct; they represent cost, latency, and computational effort. As AI systems move from experimentation to large-scale production, token consumption grows exponentially. What starts as a manageable expense during pilot phases often becomes a significant operational cost at scale. This shift is forcing enterprises to rethink how they measure value from AI.

The traditional approach has been to maximize AI usage—more prompts, more automation, more outputs. But leading organizations are now recognizing that volume does not equal value. Instead, the focus is shifting toward a more meaningful metric: outcomes achieved per unit of token consumption. In other words, how much business impact is generated for every token processed.

A major driver of inefficiency is context bloat. Many AI workflows send large volumes of unnecessary or repetitive information to models, assuming that more context leads to better results. In practice, this often has the opposite effect. Excessive context increases cost, slows down response times, and can even dilute the model’s ability to focus on relevant information. Similarly, poorly orchestrated workflows—such as redundant retries, recursive loops, or overuse of advanced models for simple tasks—further amplify token waste.

To address these challenges, forward-looking engineering teams are adopting token-aware design principles. This includes compressing and structuring context so that only relevant information is processed, dynamically selecting models based on task complexity, and instrumenting systems to monitor token consumption in real time. These approaches ensure that AI systems remain both performant and cost-effective as they scale.

Token efficiency also has broader implications beyond cost. It improves system responsiveness, enhances accuracy by reducing noise, and strengthens data security by minimizing unnecessary exposure of information. Most importantly, it enables scalability—allowing organizations to serve more users and workloads without a proportional increase in infrastructure spend.

Ultimately, token optimization is evolving into a discipline in its own right, much like financial operations (FinOps) did for cloud computing. Enterprises that embed token efficiency into their AI architecture and governance models will be better positioned to scale sustainably, control costs, and deliver measurable business outcomes. Those that do not may find that the true challenge of AI is not intelligence—but efficiency.

Agentic AI and Modern Applications – The Tip of the Iceberg

 Agentic AI represents the next evolution of Modern Applications, moving beyond passive systems of record to active, self-directed systems of action. At the surface—or the “tip of the iceberg”—organizations see copilots, chat interfaces, and task automation. But beneath that lies a deeper transformation: applications that can reason, plan, orchestrate workflows, and continuously learn from outcomes.

From a business perspective, Agentic AI shifts the narrative from “modernization as cost optimization” to “applications as intelligent business operators.” Enterprises are no longer just rewriting legacy systems—they are embedding AI agents that drive customer engagement, optimize operations, and unlock new revenue streams. This reframes deals around measurable business outcomes, not just technical upgrades.

From a technical standpoint, Modern Apps become composable, event-driven, and API-first platforms capable of hosting and coordinating AI agents. These agents integrate with enterprise data, enforce governance, and leverage cloud-native and hyperscaler AI services to act autonomously within defined guardrails.

The real opportunity lies below the surface: orchestrated ecosystems of agents collaborating across business domains. This is where competitive differentiation emerges—turning applications into adaptive, intelligent enterprises at scale.

Why Token Efficiency Matters

 Most discussions around AI tend to emphasize model capabilities—larger models, better reasoning, and more advanced outputs. However, an often overlooked yet critical factor is efficiency, particularly in how tokens are used. Token efficiency directly impacts the cost, speed, and quality of AI interactions.

First, efficient use of tokens leads to lower costs. Since most LLM platforms charge based on the number of tokens processed, minimizing unnecessary context helps reduce expenses significantly over time. Equally important is faster response time. Smaller, well-structured prompts reduce inference latency, enabling quicker interactions and better user experience.

Token efficiency also contributes to improved accuracy. By eliminating irrelevant or redundant information, the model can focus more precisely on the core query, leading to clearer and more reliable outputs. In addition, it enhances security and privacy, as sharing only essential information reduces the risk of exposing sensitive data.

Finally, efficiency enables greater scalability. Systems that optimize token usage can support more users and higher workloads without proportional increases in cost or performance bottlenecks.

In essence, token efficiency is not just a technical optimization—it is a strategic advantage for building scalable, cost-effective, and high-performing AI systems.