At IBM Client Engineering, I worked between IBM Research and client delivery. The normal path in a company that size is research, labs, product development, and then sales. Our team often cut across that. We would get on calls with research teams, work out what was actually worth taking forward, and turn it into something useful for a client.
That meant I was not just writing model code. A big part of the work was figuring out whether an idea actually held up, how to validate it, how to benchmark it properly, and how to turn it into something that made sense both technically and commercially. The work ranged from fine-tuning small language models, to climate prediction for insurance, to source-attribution work for post-training LLM behaviour.
Most of the job was taking research seriously enough to test it properly, and practical enough to make it useful.
What I Did
For every use case, we had to figure out:
- whether the underlying idea actually made sense
- how to validate it properly
- how to quantify trade-offs, limitations, and expected gains
- how to benchmark models and approaches instead of relying on vague impressions
- how to turn the result into something usable in a real business setting
Selected Outcomes
- Built RAG systems for enterprise workflows, including form-filling and document-processing use cases, with retrieval accuracy in the 75 to 85% range.
- Helped reduce end-to-end document-processing time by up to 80% by improving retrieval quality and workflow design.
- Fine-tuned language models with real and synthetic data, improving response alignment by roughly 25 to 35% in expert evaluation.
- Built large-scale content standardisation pipelines using parallel LLM execution and PySpark, reducing runtime by 97% while cutting compliance issues.
- Built MLOps workflows for monitoring, automated retraining, and lifecycle management, with drift detection reaching about 90% accuracy across large Hadoop-based environments.
- Worked on use cases outside the usual LLM demo cycle, including climate prediction for insurance and source-attribution work for post-training LLM behaviour.
Systems Built
- Built multi-agent systems using LangGraph, CrewAI, and custom tool-calling agents.
- Exposed orchestration patterns as Model Context Protocol (MCP) servers using FastAPI and async processing.
- Built natural-language-to-SQL systems so non-technical users could query complex relational data more directly.
- Built validation matrices, benchmarking flows, and observability around model behaviour so projects could be judged properly rather than by anecdote.
Client Engagements
- Fiserv: data science, data engineering, LLM fine-tuning, and LLMOps.
- Visa: data engineering, MLOps, and data science workflows.
- Sony Group: RAG systems and AI software engineering.
- USAA: Java-to-COBOL conversion workflows.
- BNY Mellon: Java-to-COBOL conversion workflows.
- Truist Financial: data engineering.
- Globant: fine-tuned LLM systems for proprietary code explanation.
Research Feedback Loop
Part of the role also involved noticing what kept going wrong across client work and feeding that back into broader research and product thinking. That meant looking at where models failed, what actually needed to be measured, and which ideas were worth carrying forward.
- Built reproducible benchmarking pipelines to judge model-task fit, generalisation, and practical usefulness.
- Designed prompts and evaluation setups for instruction-tuned and chat-based models in business-critical NLP tasks.
- Used supervised fine-tuning with LoRA where it made sense for client-specific domains.
- Compared zero-shot, few-shot, and fine-tuned approaches using both metrics and human review.
- Worked on RAG setups that combined retrieval with generative models in a way that was easier to evaluate and defend.
IBM Platform and Stack
IBM Products
- watsonx.ai
- watsonx.data
- watsonx.governance
- InstructLab
- Carbon Design System
- watsonx Code Assistant for Z
- watsonx Code Assistant for Ansible
- Ansible
Open Source and Engineering Stack
- PyTorch
- Transformers
- Scikit-learn
- Pandas
- Matplotlib
- SciPy
- Hadoop
- Hive
- PySpark
- Kerberos
- Docker
- Kubernetes
- Java
- COBOL with basic exposure