About
I am an Applied AI and Machine Learning Engineer who approaches software architecture with a first-principles mindset. My core focus is building the intelligence and distributed systems layer that connects deep learning models to mission-critical enterprise workflows serving 1M+ active ERP users across 120+ global enterprises.
Over the past 2 years, I have architected custom agent execution runtimes in Python & FastAPI, scaled Model Context Protocol (MCP) ecosystems to 1,000+ tool architectures, designed hybrid lexical+vector retrieval engines, and fine-tuned dense embedding representations (PyTorch & E5 Bi-Encoders).
In parallel, I am an active contributor to open-source AI infrastructure projects including vLLM (inference optimizations) and FastMCP / Model Context Protocol tooling. I build with an uncompromising emphasis on deterministic evaluation, zero hallucination tolerance, and sub-second execution latency.
Experience
-
2024 — PRESENT • SAP • 1M+ Active Users • 120+ Global EnterprisesArchitected and delivered the core agentic runtime, MCP tool routing layer, and evaluation control plane operating across massive enterprise ERP and talent intelligence databases (SAP HANA / SQL).
• Master Agent Harness & Planner-ReAct Runtime: Built a vendor-agnostic agent framework in Python & FastAPI with explicit Planner → ReAct → Observer lifecycle, supporting isolated child sub-agents, adaptive tool failure recovery, and circuit breakers.• Scalable MCP Catalog & Streamable HTTP: Architected a massive 1,000+ tool MCP catalog, exposing secure ERP data via FastMCP and Streamable HTTP at/lxp/mcpwith real-time SSE token and tool telemetry.• Low-Latency Tool Allowlisting: Scaled tool discovery beyond raw LLM context limits by shrinking 50 retrieved candidate tools to an optimal 20-tool allowlist using hybrid Vector embeddings + BM25F + TF-IDF + Reciprocal Rank Fusion (RRF).• Prompt Lab (Chaos Red-Teaming & Evaluation Plane): Designed a run-centric evaluation platform scoring live agent trajectories across 6 dimensions with LLM-as-Judge (0.75 pass threshold), automated red-teaming, and Golden-Set curation for DPO/RLHF pipelines.• Enterprise Hierarchical RAG: Engineered 4-level ingestion (paragraph, page, chapter, heading) over 1,000+ books with PGVector and paragraph hash deduplication.• Multilingual PII Token Classification: Built BIO-tagged entity classification models over AI4Bharat Naamapadam, WikiNER, and CoNLL datasets to redact sensitive identifiers before LLM ingestion.• E5 Bi-Encoder Fine-Tuning: Fine-tuned 335M parameter E5-large-v2 embeddings using triplet loss (anchor + positive + hard negatives) for domain competency matching with FAISS vector search.- Python
- Hierarchical RAG
- FastAPI
- Java / Spring AI
- PyTorch / E5
- MCP Protocol
- SAP HANA / PGVector
- BM25F + RRF
- SSE Streaming
-
2023 — 2024 • Clinical Intelligence • Multilingual Speech • Dense Neural RetrievalConducted applied AI/ML research and production prototyping combining quantized language models, speech processing architectures, and custom embedding fine-tuning.
• Clinical Data Intelligence: Designed progressive schema → stats → rows → summary pipelines across 116K+ patient/hospital records using local 4-bit/8-bit quantized Llama 3 models (llama.cpp) and specialized medical NER.• Indian Multilingual Speech Systems: Prototyped low-latency ASR and TTS pipelines for Gujarati and Hindi incorporating AI4Bharat open-source neural acoustic models.• Semantic Search & Microservices: Engineered asynchronous FastAPI microservices with dense vector indexing and FAISS retrieval achieving sub-100ms response latencies.- PyTorch
- Hugging Face
- llama.cpp
- Speech ASR/TTS
- FAISS
- FastAPI
Architecture & Projects
-
Master Agent
Vendor-agnostic agent execution harness built with Python & FastAPI. Features dynamic tactical tool selection (exposing top 20 tools out of 1,000+ catalog), parent-child context isolation, adaptive tool memory, circuit breakers, and real-time SSE telemetry.
- Python
- FastAPI
- Java / Spring Boot
- MCP Protocol
- Planner-Observer
- SSE Streaming
-
Hierarchical RAG
Multi-level document intelligence platform extracting paragraph, page, chapter, and heading representations across 1,000+ books. Built with GROBID, pdfplumber, PGVector, and paragraph-level hash signatures for version-aware deduplication.
- PostgreSQL / PGVector
- GROBID / OCR
- Python
- Hybrid RAG
-
Eval Control
Run-centric evaluation infrastructure scoring agent trajectories across groundedness, trajectory quality, and tool Recall@K via LLM-as-Judge with automated Chaos-Monkey style red-teaming and DPO/RLHF Golden Set curation.
- LLM-as-Judge
- Red-Teaming Harness
- FastAPI
- DPO / RLHF Pipeline
-
E5 Embeddings
Fine-tuned 335M parameter E5-large-v2 foundation model on triplet relationships (anchor + positive + hard negatives) for domain semantic retrieval, mapping millions of occupational competencies with FAISS indexing.
- PyTorch
- E5-large-v2
- FAISS Vector Search
- Hugging Face
Open Source Contributions
-
vLLM EngineCore Open Source Contributor • Inference & Memory Optimization
Active open-source contributor to vLLM, focusing on high-throughput serving optimizations, PagedAttention memory mechanics, kernel-level efficiencies, and structured output integrations.
- Python
- PyTorch / CUDA
- PagedAttention
- Inference Serving
-
FastMCPOpen Source Contributor • Python AI Tooling & Standards
Contributed to Python-based Model Context Protocol (MCP) ecosystems, standardizing asynchronous tool execution schemas, SSE streamable transports, and low-latency client integrations for agentic workflows.
- Python 3.12
- FastMCP
- AsyncIO
- JSON-RPC 2.0