AI engineering for real-world problems

Not AI for its own sake. Engineering the right use of AI.

I investigate where AI can make a meaningful difference, design how it should work with your systems and people, and build or validate the critical parts. The starting point is always the problem—not a model, framework or agent trend.

Discuss an AI problem ↗

The problems, not the buzzwords

AI has to fit the work, the systems and the consequences.

Useful AI is rarely just a prompt connected to a model. It may need to understand existing data, interact with business applications, respect permissions, handle incomplete information and know when to stop and ask a person.

I work across several problem spaces, choosing conventional software, AI or a combination based on the task and its risks.

Business workflows: reduce repetitive knowledge and document work; connect AI assistance to the applications people already use.
Software engineering: help teams understand codebases, automate bounded development tasks and improve engineering workflows.
Knowledge and information: make internal documents and fragmented information useful within a defined task or decision process.
Complex applications: explore AI capabilities within ERP, fintech and enterprise systems where reliability and control matter.
Task-specific AI: assess model adaptation, retrieval, tool use and multi-step orchestration when the problem calls for them.

A practical principle

Use the simplest architecture that meets the need.

Not every workflow needs an LLM.
Not every AI workflow needs an agent.
Not every knowledge problem needs RAG.
Not every prototype is ready for production.
Automation should have explicit permissions and failure paths.

How I approach an AI problem

From uncertainty to a system that can be trusted.

My role is to personally investigate the use case, understand its technical and operational context, and make the design decisions visible. A prototype is useful only if it tests the assumptions that matter.

Investigate the workflow

Understand users, inputs, existing systems, exceptions, current effort and the actual business outcome.

Choose the right approach

Compare deterministic automation, conventional software and AI. Define what the model should—and should not—do.

Architect the controls

Set data boundaries, permissions, tool access, validation, audit trails, fallbacks and human approval points.

Prototype and validate

Build a focused proof of concept or critical component, test it against representative cases and assess operational readiness.

Areas I can explore and build

One practice across connected problem spaces.

The right scope depends on the problem. These are directions for investigation and engineering—not a claim that every possible solution is an off-the-shelf offering.

AI-enabled business systemsWorkflow assistants, ERP/application integration, document-heavy processes and operational decision support.

AI for software engineeringCodebase-aware tooling, developer assistants, bounded agent execution and human-reviewed engineering automation.

Knowledge and task automationRetrieval and knowledge workflows, meeting assistance, task-specific agents and model adaptation where justified.

Relevant hands-on experience

Experience that informs new work.

My work and experimentation include an AI meeting assistant, Qwen3 fine-tuning, RAG and agent-based workflows, and a custom AI-native Frappe development operator built using the OpenHands SDK with human-in-the-loop approvals. I bring those lessons to new problems while validating the right design for each context.

Engineering workflows

AI-assisted Frappe development operator with bounded execution and human approval.

Knowledge workflows

AI meeting assistant and retrieval-based approaches for working with information.

Model engineering

Qwen3 fine-tuning and experimentation with model and agent workflows.