Fine-tuning whereevaluationis the work.

Fine-tuning large language models with emphasis on instruction tuning and domain adaptation. I run the loop end to end: dataset engineering, training, evaluation metrics, performance tuning and iteration on real output.

500GB+

dataset built for training and eval

25+

classifiers used as an eval suite

Top 25

Riot x AWS hackathon globally

About me

I am Vansh Bordia, an LLM model developer and subject matter expert on the LLM systems at RiftWatch. I own the full loop: dataset engineering, instruction tuning, domain adaptation, evaluation metrics and iteration driven by real model output rather than intuition.

The discipline I bring is measurement. I fine-tuned Gemma on a domain corpus, measured the results honestly, and found the real bottleneck was 500GB+ of raw data needing preprocessing and enrichment, not the model, so I built that pipeline and re-ran the comparison. I also benchmarked a managed RAG product and rejected it on cost. Mentoring is part of the job: I have coached 5,000+ students and led developer teams to national podiums.

Why me

I build domain-adapted models and the evaluation harness that proves they work: instruction tuning on curated corpora, synthetic examples for edge scenarios, classifier suites as regression checks, and continuous retraining loops so relevance does not decay as the domain moves.

LLM engineer specializing in instruction fine-tuning and domain adaptation. I fine-tuned the Gemma model family on a custom domain corpus covering terminology, strategy and coaching workflows, generated synthetic training examples for domain-specific scenarios, and paired it with RAG plus live retrieval so relevance survived a shifting domain. I am the subject matter expert on LLM systems, and I benchmark everything rather than assuming it works.

Instruction fine-tuned the Gemma model family on a custom domain dataset, with synthetic training examples for domain-specific scenarios

Built a 500GB+ preprocessing pipeline into PostgreSQL and a NoSQL store, then proved by measurement it was the real accuracy bottleneck

Developed 25+ classifiers as a regression and evaluation suite gating every model output before release

Cross-functional work across data, service and product layers, from Riot Games x AWS hackathon to client delivery for 10+ SMB clients

Coached 5,000+ students and led developer teams, building capability rather than dependency

Role stack

Large Language Models / Instruction Fine-Tuning / Domain Adaptation / PEFT / LoRA / Synthetic Data Generation / NLP Techniques / Tokenization & Embeddings / Evaluation Metrics / Performance Tuning / Benchmark Comparison / Data Preprocessing & Enrichment / Classifier Development / Hallucination Mitigation / Python / PyTorch / TensorFlow / Scikit-Learn / MLflow / RAG with Live Retrieval / Agentic AI / Prompt Engineering / Output Validation & Gating / Model Routing / MCP Servers / Gemini / Claude / Mistral / LLaMA / AWS Bedrock / Gemma / PostgreSQL / ClickHouse / BigQuery / Redis / Apache Spark