Shubham – Python, LLM, AI agent development

An internal developer harness that standardizes how engineering teams deliver software with AI coding agents. Developers use it to build and deploy software through agents while staying within pre-approved enterprise architecture and security patterns. It is designed to be token-efficient and self-improving, and it is integrated into the existing delivery workflow.

• Designed and built the harness, encoding pre-approved enterprise patterns so agent-generated code followed company standards by default.
• Implemented self-improving feedback loops so the harness refined its guidance based on real delivery outcomes.
• Optimised context and prompt usage, reducing token consumption across software delivery by ~30%.
• Integrated the harness into GitHub-based development and deployment workflows.
• Rolled it out to the internal developer community, with 100% positive feedback from adopting engineers.

An ML infrastructure platform for engineering teams across a large multi-account AWS environment. Teams use it to securely package, deploy, and serve open-source ML and generative AI models. It standardised GPU-backed deployment, lifecycle management, and governance across 100+ AWS accounts.

• Designed and implemented a secure ML serving platform spanning 100+ AWS accounts, with centralized governance and distributed deployment per business team.
• Built GPU-backed inference infrastructure on Amazon SageMaker for compute-heavy embedding and generative AI workloads.
• Developed automated packaging and deployment pipelines for Hugging Face models, reducing the manual work of moving models from experimentation to managed inference.
• Standardized model-serving patterns for deployment, networking, access control, and lifecycle management.
• Delivered reusable infrastructure templates (CloudFormation) so AI teams could deploy models consistently instead of each team building its own ML infrastructure.

A shared generative AI platform that provides reusable retrieval, agent, and model capabilities to multiple business applications. It processed 500GB+ of enterprise content and served production-grade search, embedding, reranking, and agent orchestration to downstream AI products.

• Architected the platform's core, supporting multiple AI applications across business domains.
• Built scalable ingestion and embedding pipelines processing 500GB+ of heterogeneous enterprise data.
• Implemented a ReAct-based agent architecture with LangGraph, letting applications reason, retrieve context, and invoke enterprise tools.
• Developed hybrid vector and keyword search with a reranking stage, improving context relevance before generation.
• Raised chatbot answer accuracy to ~95% while reducing response latency.
• Packaged retrieval, agent, and governance components as reusable building blocks, so new AI apps adopted common patterns instead of rebuilding them.

A production AI assistant that helps procurement teams get accurate answers from large volumes of internal documentation and enterprise data. It combines RAG, agentic workflows, and vector search with enterprise integrations to reduce reliance on manual support channels.

• Ran discovery workshops with procurement stakeholders who had low AI literacy, clarifying security, scale, and cost requirements before scoping the build.
• Architected and productionised the RAG application with LangGraph, ChatKit, and PostgreSQL/pgvector.
• Designed retrieval and grounding workflows for complex procurement documents and knowledge sources.
• Integrated MCP-based enterprise services so agents could pull context from and act on internal systems.
• Implemented orchestration, prompt management, retrieval logic, and guardrails for enterprise production use.
• Reduced inbound procurement support queries by ~60% after adoption.