// Content data — updated from Rohan's May 2026 resume const profile = { name: "Rohan Waghmare", role: "Software Developer (AI)", location: "San Francisco Bay Area, CA", status: "open-to-opportunities", summary: "I build production GenAI systems: agentic pipelines, RAG architectures, LLM fine-tuning, and MLOps on GCP, AWS, and SAP BTP. 1+ years shipping enterprise AI at scale with measurable impact. Open to relocation. Available immediately.", email: "waghmarerohan30@gmail.com", phone: "+1 (607) 245-6001", site: "rohanwaghmare.com", linkedin: "/in/rohanwaghmare", github: "/ron103", }; const experience = [ { id: "sap", company: "SAP", role: "Software Developer iXP (AI)", location: "Palo Alto, CA", start: "Jan 2026", end: "Present", period: "2026", tags: ["AI/ML", "MLOps", "Backend", "Cloud"], bullets: [ "Fine-tuned DeepSeek-14B with dual QLoRA adapters on SAP BTP via KServe + FastAPI; 98% accuracy, p95 <11s.", "Replaced a multi-hop GPT-4 pipeline at 10× lower cost via MLOps loop (Docker, ArgoCD, S3) and dynamic max_new_tokens re-inference; cut pipeline time from 1.5 min to under 30s.", "Built a RAG pipeline over a Knowledge Graph with confidence-scored retrieval; cut query-to-docs time by 92% (1hr+ to under 5 mins) for 5,000+ content developers. Exposed the system as an MCP server (SSE, JSON-RPC 2.0).", "Engineered a multi-step agentic pipeline (LangGraph + GPT-4) on SAP BTP for document extraction and backend execution.", "Extended to a production MCP server with PKCE + SAP IDP auth, audit logging, and tool-call agent observability.", ], }, { id: "mihin", company: "Michigan Health Information Network", role: "Software Engineer", location: "Binghamton, NY (Remote)", start: "Mar 2025", end: "Jan 2026", period: "2025-2026", tags: ["Backend", "DevOps", "Cloud"], bullets: [ "Built a Python data pipeline processing 500K+ records across 10+ orgs on AWS (ECS Fargate, S3, EventBridge, DynamoDB) with Docker/Terraform and cross-month aggregation.", "Cut processing latency 40% via S3 Select API with conditional ScanRange and multipart parallelization (ThreadPoolExecutor, 20MB chunks) across multi-GB files.", ], }, ]; const projects = [ { id: "deepseek-pipeline", title: "DeepSeek Fine-Tuning Pipeline on SAP BTP", year: "2026", kind: "MLOps", tags: ["QLoRA", "DeepSeek", "KServe", "ArgoCD", "AWS"], blurb: "Swapped a multi-hop GPT-4 pipeline for a fine-tuned DeepSeek-14B model with a full MLOps loop. 10× cheaper, 98% accurate.", metrics: [ ["98%", "domain extraction accuracy"], ["10×", "cost reduction vs GPT-4"], ["p95 <11s", "inference"], ], details: { problem: "The incumbent domain-extraction pipeline chained multiple GPT-4 calls: slow, expensive, hard to reproduce. Each hop added latency and cost.", approach: [ "Fine-tuned DeepSeek-14B with dual QLoRA adapters on curated SAP domain data, versioned in HuggingFace.", "Built the full MLOps loop: ArgoCD for GitOps, KServe for model serving, FastAPI for the inference gateway, Docker + AWS S3 for artifact storage.", "Used dynamic max_new_tokens re-inference to cut pipeline time from 1.5 min to under 30s with zero downtime.", ], stack: ["QLoRA", "DeepSeek-14B", "HuggingFace", "KServe", "ArgoCD", "FastAPI", "Docker", "AWS S3", "SAP BTP"], outcomes: [ "98% domain extraction accuracy on the SAP eval set.", "p95 <11s inference; pipeline time cut from 1.5 min to under 30s.", "10× lower cost than the GPT-4 chain it replaced.", ], }, }, { id: "mcp-rag-server", title: "RAG + MCP Server for SAP HANA", year: "2026", kind: "Developer Tooling", tags: ["MCP", "RAG", "SPARQL", "LangGraph", "SSE"], blurb: "Confidence-scored RAG over SAP's Knowledge Graph, exposed as an MCP server. Query-to-doc: 1hr → 5min for 5,000+ content developers.", metrics: [ ["92%", "time reduction"], ["5,000+", "engineers served"], ], details: { problem: "SAP developers spent up to an hour stitching together SPARQL queries and documentation for every non-trivial question. The process was manual, error-prone, and blocked by auth.", approach: [ "Built a RAG pipeline over SAP's SPARQL Knowledge Graph with confidence-scored retrieval; low-confidence results trigger automatic query refinement.", "Exposed it as an MCP server with SSE endpoints and JSON-RPC 2.0 bidirectional streaming, making Claude and Cline first-class clients.", "Implemented enterprise-grade auth/auditing through PKCE + SAP IDP, so the server works inside SAP's security perimeter.", "LangGraph orchestrates multi-step agentic sequences when a single retrieval isn't enough.", ], stack: ["MCP", "SSE", "JSON-RPC 2.0", "SPARQL", "RAG", "SAP HANA", "LangGraph", "FastAPI", "PKCE"], outcomes: [ "Query-to-documentation time: 1hr → 5min (−92%).", "5,000+ content developers using it in production.", "No extra API keys needed. Auth flows through PKCE + SAP IDP.", ], }, }, { id: "layoff-tracker", title: "Industry-Specific Layoff Tracker", year: "2024", kind: "Data Pipeline", tags: ["Python", "Flask", "MongoDB", "NLP"], blurb: "Real-time ingestion of 208K+ Reddit & 4chan posts/month with sentiment + toxicity scoring and Plotly dashboards.", metrics: [ ["208K+", "posts / month"], ["98%", "sentiment accuracy"], ], details: { problem: "Layoff conversation is fragmented across forums. How do you surface early, industry-specific signals before they hit the news?", approach: [ "Built a Faktory-backed ingestion pipeline crawling Reddit + 4chan for posts matching industry + role vocabularies.", "Ran each post through NLTK + a fine-tuned classifier for sentiment and toxicity scoring, stored in MongoDB.", "Exposed Python/Flask REST APIs powering Plotly dashboards for trend visualization by time, company, and industry.", ], stack: ["Python", "Flask", "MongoDB", "Faktory", "NLTK", "Plotly", "REST"], outcomes: [ "98% sentiment/toxicity classification accuracy on labeled set.", "Ingests 208K+ posts/month at steady state.", ], }, }, { id: "tb-detection", title: "Tuberculosis Detection via Transfer Learning", year: "2023", kind: "Research / Medical AI", tags: ["TensorFlow", "ResNet-50", "Computer Vision", "IEEE"], blurb: "ResNet-50 transfer learning on TBX11K, validated with Grad-CAM / F1 / AUC. Co-authored an IEEE paper.", metrics: [ ["92%", "accuracy"], ["60+", "studies reviewed"], ], details: { problem: "Build a reproducible medical AI pipeline for chest X-ray TB detection, with interpretability as a first-class concern.", approach: [ "Trained ResNet-50 in TensorFlow on TBX11K with CLAHE preprocessing and augmentation.", "Validated predictions with Grad-CAM heatmaps, F1, and AUC so clinicians can inspect why the model fired.", "Reviewed 60+ studies to situate the work; co-authored an IEEE-published paper on the pipeline.", ], stack: ["TensorFlow", "ResNet-50", "Python", "Grad-CAM", "CLAHE"], outcomes: [ "92% accuracy on TBX11K test split.", "Grad-CAM overlays line up with clinically salient regions.", "Published IEEE paper.", ], }, }, ]; const education = [ { school: "Binghamton University, SUNY", degree: "M.S. Computer Science", period: "Aug 2023 - Dec 2025", courses: "Algorithms · Operating Systems · Computer Networks · Design Patterns · Data Science Pipeline", }, { school: "School of Engineering, MIT ADT", degree: "B.Tech Computer Science", period: "Aug 2019 - May 2023", courses: "Databases · Machine Learning · Deep Learning · Operations Research · Big Data", }, ]; const skills = [ { group: "Languages", items: ["Python", "JavaScript", "TypeScript", "React", "Node", "Express", "C", "SQL"], }, { group: "AI / ML", items: ["LLM Fine-tuning (QLoRA)", "HuggingFace", "MCP", "LangGraph", "Agentic", "TensorFlow", "Scikit-learn", "RAG", "OpenCV"], }, { group: "Cloud & DevOps", items: ["GCP", "Vertex AI", "BigQuery", "Cloud Run", "Pub/Sub", "Vector Search", "AWS", "SAP BTP/AI Core", "AWS CCP"], }, { group: "Backend & APIs", items: ["FastAPI", "Flask", "Django", "REST", "MCP", "SSE", "JSON-RPC 2.0", "SPARQL", "RDF"], }, { group: "Databases", items: ["PostgreSQL", "SAP HANA", "MongoDB", "Vector DB (Pinecone)"], }, { group: "Tools", items: ["Git", "Linux/Unix", "Docker", "Terraform", "Postman/Bruno", "Pandas", "NumPy", "Grafana", "Agile/SCRUM"], }, ]; window.__data = { profile, experience, projects, education, skills };