Headline

ML Engineer · LLMOps & MLOps · LLM Evaluation Platforms · RAG · Fine-tuning · sub-50ms Inference · Open to Remote, Hybrid

About

Most companies pay $500–$5,000/month for LLM observability. I built it from scratch. On free GPUs. In 4 days. ━━━━━━━━━━━━━━━━━━━━━━ I'm Pranav — an ML Engineer who doesn't stop at the notebook. I build production ML systems end-to-end: from model training and fine-tuning to LLM evaluation pipelines, drift detection, inference optimization, and live deployment with real performance numbers. Here's what I've shipped: → LLM-Eval Platform — dual AI-judge pipeline (Phi-3 + Claude API) scoring factuality, coherence, toxicity, instruction-following · 94% inter-rater agreement · zero schema violations across 500+ API calls · auto LoRA fine-tuning on drift · 8-service Docker Compose · HuggingFace Spaces → BERT → ONNX → INT8 — inference pipeline benchmarked across 3 precisions · p50/p95/p99 latency table · Locust 100-user load test · cost-per-request on AWS ml.m5.xlarge · sub-42ms p50 → DistilBERT MLOps — 92.4% accuracy · F1 0.924 · Feast feature store sub-5ms retrieval · ~$3K/month compute saving at 1K req/s · Nginx canary deployment · 100+ HuggingFace downloads → Llama-3-8B LoRA — DeepSpeed ZeRO-2 on Kaggle 2×T4 · ~40% VRAM reduction · W&B public dashboard · PEFT adapters on HuggingFace Hub → RAG Pipeline — LangChain + Claude 3.5 Sonnet · Hit@3: 0.91 · MRR: 0.87 · BM25 + dense hybrid retrieval · 200-query benchmark ━━━━━━━━━━━━━━━━━━━━━━ My stack: PyTorch · HuggingFace · vLLM · FastAPI · LangChain · Anthropic Claude API · OpenAI API · Evidently AI · MLflow · DVC · Docker · GitHub Actions · AWS (EC2, S3, Lambda, SageMaker) · Streamlit · ONNX · TensorRT · LoRA · PEFT · DeepSpeed · W&B · DagsHub · OpenRouter · Phi-3 · Mistral · Llama · Kimi K2.6 ━━━━━━━━━━━━━━━━━━━━━━ Currently open to: → ML Engineer · LLMOps Engineer · MLOps Engineer · GenAI Engineer → Remote-first · India · Hybrid · Relocation → Full-time · Internship · Contract Everything I build is open source: 🔗 github.com/shreecloud-ai 🤗 huggingface.co/shreecloud 📊 Public W&B dashboards · DagsHub MLflow · Medium articles If you're building something in the LLM eval, inference optimization, or MLOps space — let's talk. pranav.shree96@gmail.com

Location

Kolhapur

Work experience

ML Engineer — Independent Projects & Open Source at Self-Employed, Sep 2025 - Present · 10 mos, Remote, Building production-grade ML systems end-to-end — LLM evaluation platforms, inference optimization pipelines, RAG systems, and distributed fine-tuning infrastructure. All projects deployed, benchmarked, and open-sourced. Key deliverables: • LLM-Eval Platform — dual AI-judge pipeline, 4-model benchmark, auto LoRA fine-tuning trigger, 8-service Docker Compose, Evidently AI drift detection, HuggingFace Spaces deployment • BERT→ONNX→INT8 Benchmark — FP32/FP16/INT8 latency table, Locust 100-user load test, cost-per-request on AWS ml.m5.xlarge, sub-42ms p50 • DistilBERT MLOps — 92.4% accuracy, Feast feature store, Nginx canary, 100+ HuggingFace downloads • Llama-3-8B LoRA — DeepSpeed ZeRO-2, Kaggle 2×T4, ~40% VRAM reduction, W&B public • RAG Assistant — LangChain + Claude API, Hit@3: 0.91, MRR: 0.87, hybrid BM25 + dense retrieval Stack: PyTorch · HuggingFace · vLLM · FastAPI · LangChain · Anthropic Claude API · Evidently AI · MLflow · Docker · AWS · ONNX · TensorRT · LoRA · PEFT · DeepSpeed · W&B

Education

Master of Computer Applications - MCA, Computer and Information Sciences and Support Services, DY Patil Technical Campus,Talsande(Autonomous)

Bca, Business Administration and Management, General, shivaji university, kolhapur