IBM Client Engineering

AI Engineer · AI & Hybrid Cloud, watsonx Client Engineering · Bangalore

30 October 2023 to 25 July 2025

At IBM Client Engineering, I worked with IBM Research and client teams to develop AI systems for business problems, which meant understanding the proposed approach, testing whether it was useful, and building the software needed to use it within the client’s environment.

Agents and access to data

• I built LLM agents and execution harnesses in raw Python and evaluated LangGraph and CrewAI, choosing the approach according to the engagement rather than committing every system to one framework.
• I productionised the systems as FastAPI and MCP services, using asynchronous parallel execution and LangSmith observability, and reused the orchestration patterns across engagements.
• I developed natural-language-to-SQL interfaces for complex relational databases, so non-technical business users could query their data more directly without depending on an analyst for every request.

Retrieval and source attribution

• I built RAG systems with hybrid vector retrieval, including an enterprise form-filling assistant that processed thousands of documents and reduced end-to-end workflow time by approximately 80%.
• I led a three-person team building a React interface that highlighted the exact PDF line behind each answer, giving users a way to inspect the source, and the interface was released as a reusable asset across IBM Client Engineering.

Model development and evaluation

• I fine-tuned enterprise-domain LLMs using LoRA-based supervised fine-tuning on real and synthetic data, and compared zero-shot, few-shot, and fine-tuned approaches through reproducible benchmarks using F1, ROUGE, and BLEU, with round-trip validation against precomputed source metrics.
• I applied layer pruning and knowledge distillation for small language model deployment, and built validation matrices and benchmarking workflows that combined quantitative metrics with human review.
• I developed end-to-end pipelines for computer vision, time-series forecasting, and tabular modelling, covering preprocessing through deployment on financial-risk data, logs and queries, satellite imagery, and environmental data. The wider work included climate prediction for insurance and source attribution concerning post-training LLM behaviour.

Code modernisation and knowledge sharing

• I was the team’s subject-matter expert for watsonx Code Assistant for Z, with the most engagements in that area, and was the only engineer building the Java validation workflow within the 15-person FSM team for the COBOL-to-Java work.
• I ran knowledge-transfer sessions for the team and a department of approximately 150 people, so colleagues could use the workflow and the experience gained through those engagements.

Monitoring and client infrastructure

• I built MLOps workflows for monitoring, automated retraining, and lifecycle management, validating drift detection through k-fold cross-validation on stratified subsets.
• I developed a custom Spark wrapper for Watson OpenScale on Cloud Pak for Data/OpenShift, which used Apache Livy to support batch monitoring of more than 300 models directly on client Hadoop and Hive clusters.

Client engagements

The responsibilities differed by client, so this table identifies the main areas of work without attributing every result above to every engagement.

Tools and skills

Working across these engagements helped me understand how research ideas change when they meet a client’s data and operational requirements, and I fed recurring model failures and evaluation needs back into research and product discussions.