NISHIKA YADAV

Hi, I am a Backend & AI Engineer with 2+ years building production LLM platforms, multi-provider orchestration, RAG pipelines, vector stores, and fine-tuning models. Python is my primary language, and I bring experience in data science, machine learning, and core AI as well.

My interest lies with a plethora of things from geospatial, remote-sensing, and satellite data to responsible AI, AI Governance, and tech/public policy. I hope to get the chance to work on these fields in my near future or dwell into them as much as I can on my own in the meantime.

• ->Architected a Python-based multi-provider LLM gateway (OpenAI, AWS Bedrock, Gemini) with a provider-agnostic abstraction layer in FastAPI + SQLModel
• ->Built a production document upload to optional document transformation to vector store pipeline with asynchronous support via Celery + Redis + gevent
• ->Contributed to a guardrails microservice with autodiscovery, A/B testing, and service-to-service auth — validators added without modifying core platform code.
• ->Delivered features on a production LLMOps frontend (Next.js 16, React 19, TypeScript): guardrails config UI, document management, knowledge base creation, speech-to-text evaluation.
• ->Engineered complete pipeline for fine-tuning LLM Models and evaluating those models with OpenAI — stratified data splits, background job tracking, JSONL preprocessing.
• ->Experimented with multiple file search setups including AWS Bedrock + OpenSearch + S3 and Google Gemini File Search to evaluate best-fit RAG setups for NGO clients.
• ->Owned the full prompt engineering and evaluation workflow for a public policy chatbot serving government stakeholders, continuously optimizing prompt templates and RAG configurations using structured evaluations and real-world user feedback.

• ->Built DBT transformation pipelines converting raw NGO field data into analytics-ready models powering Jal Jeevan Mission dashboards.
• ->Delivered Apache Superset dashboards tracking key mission metrics, replacing manual reporting workflows.

• ->Delivered a production chatbot (Ollama + LlamaIndex + LangChain) letting customers query their cash flow database in natural language, replacing a 3-step manual workflow.
• ->Fine-tuned Llama 2 and CodeLlama for Text-to-SQL using SFTTrainer, ORPOTrainer, DPOTrainer; explored QLoRA 4-bit quantisation to run 7B models on constrained hardware.

• ->Produced daily analysis reports in Python and Tableau, surfacing trends in Tamil Nadu's disability data for the research team.

• ->Trained UX-Net, a modified U-Net architecture combining convolutional and recurrent processing for low-latency speech separation, and applied it to speech separation and acoustic echo cancellation tasks in PyTorch.

Tech Volunteer

Volunteer