Senior AI Engineer working on LLM-centric systems, including conversational platforms, agentic workflows, and retrieval-augmented generation (RAG) pipelines, with an emphasis on production reliability, inference efficiency, and system design.
Portfolio: https://shaileshkumar97.github.io/ (under maintenance)
I design and build end-to-end GenAI systems that operate under real production constraints - latency, cost, reliability, evaluation, and observability.
My work sits at the intersection of:
- LLM inference & orchestration
- Agentic and multi-agent system design
- RAG pipelines and knowledge-grounded reasoning
- ML infrastructure & MLOps
I care less about demos and more about systems that hold up in production.
-
LLM-powered conversational platforms
Architecting assistants that integrate structured data, unstructured knowledge, and tool usage. -
Agentic & Multi-Agent Systems
Task decomposition, planning, tool execution, memory, and coordination between agents. -
Retrieval-Augmented Generation (RAG)
Indexing strategies, chunking trade-offs, retrieval quality, grounding, and hallucination mitigation. -
Inference & System Efficiency
Latency optimization, batching, prompt optimization, cost-aware deployments. -
Evaluation & Monitoring
Offline/online evaluation, drift detection, quality metrics, and feedback loops for LLM systems.
- Prompting & prompt orchestration
- Tool calling & function execution
- Context management & memory
- Multi-step reasoning pipelines
- Agent planning & control loops
- Multi-agent coordination patterns
- State, memory, and tool abstractions
- Embedding strategies
- Vector search & hybrid retrieval
- Query rewriting & reranking
- Knowledge freshness & consistency
- Feature pipelines & stores
- Model lifecycle management
- CI/CD for ML & GenAI systems
- Observability, monitoring, and drift detection
Low-latency, real-time conversational voice agent
- Chained architecture: STT -> LLM -> TTS for streaming voice interaction
- Token-level streaming to reduce response latency and support interruptions
- Agent logic designed for predictable timing and conversational turn-taking
- Clear separation between backend agent orchestration and frontend audio handling
- Built with production-style constraints around latency and reliability
Try it here (Free Tier APIs): https://huggingface.co/spaces/Shailesh-Kumar/Appointment-Voice-Agent-Frontend
Backend: https://github.com/ShaileshKumar97/Appointment-Voice-Agent-Backend
Frontend: https://github.com/ShaileshKumar97/Appointment-Voice-Agent-Frontend
Conversational planning assistant using LangGraph & Streamlit
- Natural-language planning with multi-turn dialogue
- Clarification prompts when requests are ambiguous
- Structured plan creation, versioning and diff visibility
- Context compression to manage token limits
- Executive summaries & persistent plan state
Try it here: https://huggingface.co/spaces/Shailesh-Kumar/Planning-Agent
https://github.com/ShaileshKumar97/Planning-Agent
Natural language interface for querying structured data
- NL → SQL with conversational context
- Schema-aware reasoning
- Designed for API-first, production usage
https://github.com/ShaileshKumar97/ChatQL
Reference implementations for scalable backend systems
- Service boundaries & API design
- Containerized deployments
- Patterns applicable to ML/AI services
https://github.com/ShaileshKumar97/Microservice-Architectures
- Vision Transformer–based ECG classification
- Speech enhancement pipelines
- Classical ML systems and experimentation workflows
- An opensource package for drift detection & data quality management
https://github.com/katonic-dev/explainit
Focused on building AI systems that work beyond the demo.