LinkedIn · Email · X · IEEE paper
I work on the unglamorous half of LLM engineering: getting systems that demo well to stay correct, fast and affordable in production. At Celebal Technologies I ship agentic document workflows for enterprise clients on Azure, AWS, GCP and Databricks. Outside work I build tools for agents: web search, memory, crawling. This page is laid out like a paper because that is how I think about the work. Claims need baselines, and results need evals.
| Agents | LangGraph supervisor graphs, MCP tools, async FastAPI |
| Retrieval | Hybrid BM25 + dense, cross-encoder rerank, GraphRAG on Neo4j |
| Post-training | QLoRA → DPO on open 7B models, served with vLLM + AWQ |
| Evals | RAGAS, LLM-as-judge, Langfuse, CI gates on faithfulness |
The GraphRAG pipeline indexes 15K+ documents and improves retrieval accuracy by about 35% over a vector-only baseline. The Text-to-SQL agent uses human-in-the-loop validation and lets 50+ non-technical users run their own analytics, so a request that used to take days now takes minutes.
flowchart LR
Q([query]) --> C{cache}
C -- miss --> H[BM25 + dense] --> R[rerank] --> G[LLM] --> A([answer])
C -- hit --> A
G -. traces .-> E[RAGAS eval gate]
E -- regression --> X[[merge blocked]]
Figure 3. Eval-gated RAG over 2,000+ SEC 10-K filings.
| GEXA | Web search API for AI agents |
| MindLayer | Switch LLMs without losing memory |
| Web-Crawler | Async crawler with JS rendering and stealth |
| Manan | Privacy-first mental health app |
[1] "Enhancing Assistive Communication through IoT-Based Hand Gesture Recognition," IEEE CCPIS, 2025. [IEEE Xplore]
Appendix A Stack
agents LangGraph, Semantic Kernel, MCP, FastAPI (async), Pydantic
retrieval FAISS, Neo4j, BM25, cross-encoders, Azure AI Search, Databricks Vector Search
training PyTorch, TRL, Unsloth, PEFT (LoRA / QLoRA), DPO, LayoutLMv3
serving vLLM, AWQ, Docker, Kubernetes, Terraform
ops MLflow, Langfuse, RAGAS, GitHub Actions, Azure DevOps
cloud Azure OpenAI, AWS (SageMaker, Bedrock, ECS), GCP Vertex AI, Databricks