Hiring for LLM Developer? Start here.
I build LLM and agent systems that survive production: agentic harnesses with tool orchestration, RAG over live data, and 25+ classifiers as a release gate on every output.
500GB+
preprocessed into 25+ classifiers
Top 25
Riot x AWS hackathon globally
20+
orgs served at sub-500ms P90
About me
I am Vansh Bordia, an LLM engineer and the subject matter expert on LLM systems at RiftWatch: fine-tuned models, the agentic harness, the retrieval layer and the 25+ classifier validation suite that gates every output. I served as sole technical owner across model, data and service layers through to acquisition.
The interesting work is the unglamorous part. I fine-tuned Gemma on a domain dataset, measured it honestly, and found the bottleneck was 500GB+ of raw data needing preprocessing, not the model, so I built that pipeline. I have also rejected a managed RAG option on cost grounds once I benchmarked it. Grounding in live data beats stale weights, and a classifier gate beats optimism.
Why me
I build LLM systems that hold up in production: fine-tuned models on domain data, an agentic harness routing models and tools, retrieval grounded in live sources, and the validation suite that decides whether any of it is allowed to ship.
LLM engineer building fine-tuned, retrieval-grounded and agentic systems in production. I fine-tuned the Gemma family on a custom domain dataset covering terminology, strategy and coaching workflows, paired it with RAG and live retrieval so answers stayed grounded as the domain shifted, and built a 500GB+ preprocessing pipeline feeding 25+ classifiers. My bias is that evaluation methodology, not model capability, is the real bottleneck.
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 it was the real accuracy bottleneck
25+ classifier validation suite gating all model output, developed across aggression, KAST, KD, KDA, headshot percentage and ability usage
Agentic harness with intent understanding, routing and tool orchestration, exposed via REST APIs and MCP servers
Top 25 globally, Riot Games x AWS hackathon; benchmarked Bedrock Knowledge Base against a custom pipeline and rejected it on cost
Type-safe Python library on PyPI adopted by 100+ developers, official Riot API access in 2 weeks
Role stack
Python / PyTorch / TensorFlow / Scikit-Learn / MLflow / Natural Language Processing / Tokenization & Embeddings / Evaluation Metrics / Performance Tuning / LoRA / PEFT / Model Quantization / Prompt Caching / LangChain / GenAI Frameworks / Agentic AI / Agentic Harness Design / Tool & Function Calling / Multi-Step Reasoning / MCP Servers / Model Routing / RAG with Live Retrieval / Hallucination Mitigation / Instruction Tuning / Domain Adaptation / Synthetic Training Data / Fine-Tuning (Gemma, LLaMA, Mistral) / Prompt Engineering / Evaluation & Evals / Output Validation & Gating / Classifier Development / Data Preprocessing & Enrichment / Structured Extraction / Vector Databases / OpenAI-Compatible APIs / Gemini / Claude / AWS Bedrock / FastAPI / PostgreSQL / ClickHouse / BigQuery / Redis / Apache Spark / ETL/ELT Design / Docker / Kubernetes / TypeScript / React
Relevant work
Esports intelligence platform, acquired by a pro org
Type-safe Python library for Valorant data, published on PyPI
Interactive SQL schema visualizer, 100% client-side
Contact
Need LLMs that survive production traffic? Talk to me.