Nupur Sharma is a Qodo solutions engineer working on a central problem in AI-assisted software development: making automated code review relevant, reliable, and responsive to the standards of individual engineering teams. Her background in DevSecOps informs her approach to agents whose reasoning, retrieval, and recommendations are less predictable than conventional delivery pipelines.
Based in the Netherlands, Sharma previously worked in solution design and development engineering at ABN AMRO. Her professional background also includes AWS certifications in development, systems operations, and solutions architecture. At Qodo, her work encompasses contextual code generation, automated testing, and pull-request review.
- Selective context engineering. Larger context windows do not guarantee better decisions: agents may overlook instructions buried within lengthy prompts or lose their objective when given entire repositories. Sharma favors selecting task-specific information through ranked retrieval, hierarchical summaries, iterative retrieval, or dependency-aware knowledge graphs. Each carries different costs in indexing, model processing, maintenance, and developer effort; critic nodes can catch drift but increase latency.
- The orchestration paradox. Powerful reasoning models can consume their budgets researching tools and reconsidering approaches without completing their assignments. Sharma describes an 80/20 hybrid agent architecture that reserves flexible reasoning for exploration while using deterministic validation, summarization, retry limits, and timeouts to produce bounded results.
- Specialized multi-agent code review. In Qodo’s architecture, a context collector gathers material from pull requests, connected tools, and contextual infrastructure, then supplies focused inputs to agents examining security, code changes, and issue-tracker requirements. A coordinating judge agent reconciles their recommendations against the specific change. LangChain supports communication between agents, with intermediate outputs refined into context for subsequent steps.
- Organizational judgment in automated review. Sharma calibrates recommendations using pull-request history, architectural guidelines, compliance rules, and developers’ responses. She distinguishes precedent from correctness: repeated insecure practices should not become acceptable simply because they appear in previous reviews.
Her AI Engineer presentation on context overload frames dependable agentic development around constrained inputs, specialized responsibilities, and explicit engineering judgment.
