SHIVANAND N.

Most people building LLM products have never trained one. That's why the fixes stop at the prompt, when retrieval quietly degrades, cost per conversation triples, or quality regresses, and nobody notices. The problem is underneath the API, and that's where I work. Five years training language models, and five years shipping systems built on them. That combination is why teams call me when what they already built stops holding up. WHAT I DO 1) Production LLM systems - RAG, agents, serving Retrieval that actually retrieves: 92% retrieval accuracy on a LlamaIndex + Weaviate pipeline with a 20% cut in query time. Chunking, embedding choice, hybrid and reranked retrieval - and an eval set that proves the change helped instead of a vibe check. Agent systems with a cost and latency budget: multi-agent pipelines in LangChain / LlamaIndex / CrewAI, tool calling, long-running state. One automated product-information system cut manual review effort ~90%. Self-hosted and open-weight serving: GPU sizing, quantisation strategy, throughput and concurrency planning, and quality validation against a frontier baseline before you cut over. At Dell I shipped 4-bit GPTQ quantisation and SparseGPT pruning (~40% sparsity) for hardware-constrained inference. Latency and cost: replacing an LLM call with a fine-tuned 300M classifier took one production path from ~1s to ~100ms. Model routing, caching, honest per-request cost accounting. Evaluation and regression gates: offline eval sets, LLM-judge calibration, CI gates so a prompt or model change can't silently regress. Most teams I meet have no way to answer "is it better than last week." 2) Fine-tuning, post-training and alignment Pre-trained a 355M-parameter GPT-2-medium architecture from scratch on 28B tokens (Cosmopedia-v2), distributed across 4x NVIDIA H100's with DeepSpeed - mixed precision, gradient accumulation, LR scheduling. Beat the original GPT-2-medium checkpoint on perplexity. Improved Phi-4-14B-Instruct by 2% across every Hugging Face leaderboard benchmark via Model Stock merging, validated cheaply first on a LoRA-tuned Qwen2.5-1.5B proxy over 1.2M curated STEM samples. LoRA-tuned Qwen2.5-14B-Instruct on 12K reasoning samples, using synthetic data from a multi-agent generation pipeline - measured gains on GSM8K, GPQA, and MMLU. LoRA + DPO on Llama-2 for customer-care summarisation (23K SFT samples, 5K preference pairs): 17% better across evaluation metrics. Designed and ablated a novel Drift-Diffusion attention mechanism on BERT-base, with full Weights & Biases tracking across baseline, unscaled and gated variants. The honest version: most projects that arrive asking for a fine-tune don't need one. The base model was already good enough, the eval set couldn't detect improvement, or the problem was retrieval. I'll tell you which before you spend GPU budget - that answer is worth more than the training run. STACK Python, PyTorch, Hugging Face, DeepSpeed, Weights & Biases, FastAPI, LangChain, LlamaIndex, CrewAI, Weaviate, Elasticsearch, MongoDB, Docker, Kubernetes, AWS (EC2, Inferentia-2), GCP. OpenAI, Anthropic, Gemma / Llama / Mistral / Qwen / Phi. ElevenLabs and LiveKit for voice. BACKGROUND ML Engineer at Dell Technologies and BYJU'S AI Labs, where a multi-objective Transformer recommender I built served 1M+ students across 1B+ data points at 85% F1. Contributor to Hugging Face Transformers documentation and to DocsGPT. Top-Rated on Upwork with a 100% Job Success Score. WHO I'M NOT FOR If the job is wiring Zapier or n8n between two SaaS tools, hire a generalist - genuinely, you'll get a better deal and a faster one. I'm worth the rate when the system has to be correct, cheap and measurable under real traffic. HOW TO START Start with the fixed-fee diagnostic rather than an open-ended hourly build. I read your pipeline, your traces and your evaluation setup, then send a written diagnosis: where quality is leaking, what each request actually costs, what to fix first, and what fixing it takes. It stands on its own as a deliverable, and it becomes the scope if you want me to do the build. Send me what's breaking and one example of the wrong output. That's enough to start.