I build end-to-end AI/ML systems spanning RAG, agentic AI, LLM fine-tuning, and applied deep learning, with a production software engineering background in building scalable APIs, data pipelines, and backend systems.

• Hybrid RAG Suite
  • Built a RAG app that ingests PDFs, chunks text, stores embeddings in Chroma, and supports chat over each uploaded document set.
  • Implemented hybrid retrieval that combines dense embeddings with BM25, merges ranks with Reciprocal Rank Fusion, and supports optional cross-encoder reranking.
  • Built a web Q&A mode that queries Tavily search API, Wikipedia, and arXiv in parallel with timeouts, then generates answers grounded in retrieved sources.
  • Created an evaluation module scoring Hit@K, Recall@K, Precision@K, MRR, MAP@K and nDCG@K from expected-text matches or gold chunk id labels, used to compare dense, BM25 and hybrid retrievers with and without reranking.
  • Added LangSmith observability: each chat request is one trace covering question rewrite, dense/BM25/RRF retrieval, cross-encoder reranking and the answer call, with RAG evaluation runs grouped into a single trace.
  • Demo: https://huggingface.co/spaces/ash001/Hybrid_RAG_Suite
  • Repo: https://github.com/sparklerz/Hybrid-RAG-Suite
• LangGraph Multi-Agent Orchestrator
  • Built a Streamlit multi-agent chat app that routes each query to SQL, Neo4j, web tools, or a general assistant using LangGraph state graphs.
  • Implemented a read-only SQLite SQL agent with LangChain tool-calling, capped at 30 steps and 60 sec per run.
  • Built a Neo4j agent that generates Cypher from the graph schema, executes it, then answers from query results.
  • Exposed the Tools Agent's web search (DuckDuckGo), Wikipedia, arXiv and calculator tools as standalone MCP servers (FastMCP, stdio); the orchestrator is the MCP client (langchain-mcp-adapters), discovering tools at startup and calling them asynchronously.
  • Added LangSmith observability: each chat turn is one trace with the router, every LLM call and MCP tool calls nested inside, and the Graph Agent's Cypher and answer calls grouped under one run.
  • Deployed to Hugging Face Spaces using GitHub Actions CI/CD.
  • Demo: https://huggingface.co/spaces/ash001/LangGraph_Multi-Agent_Orchestrator
  • Repo: https://github.com/sparklerz/LangGraph-Multi-Agent-Orchestrator
• Gemma-2B QLoRA Adapter Fine-Tuning
  • Fine-tuned google/gemma-2b with QLoRA on Databricks Dolly-15k, using 2,000 train and 200 eval samples.
  • Trained with 4-bit NF4 (double quant), gradient checkpointing, and 4-step accumulation.
  • Trained LoRA adapters with r=8, alpha=16, dropout=0.05 on attention and MLP projection layers.
  • Logged metrics to W&B, saved adapter weights with a run_metadata.json, and published a base vs tuned comparison report.
  • Adapter: https://huggingface.co/ash001/gemma-2b-dolly-qlora-adapter
  • W&B report: https://wandb.ai/kannansarat9/gemma-qlora
  • Repo: https://github.com/sparklerz/Gemma-2B-QLoRA-Adapter-Fine-Tuning
• Deep Learning Fundamentals Suite
  • Built a multi-page Streamlit app with 6 deep learning demos across tabular data, NLP, time series, and image classification.
  • Packaged each model for repeatable inference by saving weights and preprocessing files, then hosting them on Hugging Face Hub for app downloads and caching.
  • Trained models in TensorFlow/Keras and PyTorch, including ANNs, SimpleRNNs, LSTMs, and ResNet-18 transfer learning, reaching 97.9% validation accuracy on Cats vs Dogs and 2.86 RMSE on NYC taxi fare.
  • Demo: https://deep-learning-fundamentals-suite.streamlit.app/
  • Repo: https://github.com/sparklerz/Deep-Learning-Fundamentals-Suite

Before focusing more broadly on AI/ML engineering, I explored distributed and decentralised LLM training and inference.

• Hivemind fine-tuning (Qwen2-0.5B) — Internet-scale data parallelism with DHT + fault tolerance; measured val-loss reductions.
  • Overview: (meta repo) https://github.com/sparklerz/hivemind-qwen2-0.5b
  • Article: https://medium.com/@kannansarat9/finetuning-qwen-0-5b-using-hivemind-data-parallelism-over-the-internet-e20af1b15c05
• Petals (LLaMA-2-70B) — Decentralised inference + Deep prompt-tuning via swarm model-parallelism.
  • Overview: (meta repo) https://github.com/sparklerz/petals-llama2-70b
  • Part 1: https://medium.com/@kannansarat9/part-1-inferencing-llama-2-70b-using-petals-swarm-model-parallelism-over-the-internet-a29de8f8aef3
  • Part 2: https://medium.com/@kannansarat9/part-2-prompt-tuning-llama-2-70b-using-petals-model-parallelism-over-the-internet-89cdee667840
• Distributed Multi-GPU LLM Fine-Tuning (monorepo) — PyTorch (DDP, FSDP), DeepSpeed (ZeRO Offload, Pipeline Parallelism), Ray (Train, Tune), MosaicML; W&B/MLflow + HF Hub for fully traceable runs.
  • Repo: https://github.com/sparklerz/multigpu-llm-finetuning

Medium: https://medium.com/@kannansarat9

DMs open: https://x.com/saratkannan