Kushal Seemakurthi

SYSTEM: ACTIVE

I architect systems,
not just models.

Senior Data Scientist & POD Lead with 4.5 years building production-scale GenAI, NLP, and Document AI systems that drive enterprise revenue.

See the work ↓

0

% of company ARR

0

documents / month

0

efficiency gain

0

revenue generated

I'm a Senior Data Scientist & POD Lead at Ushur, where I lead a 12-engineer cross-functional POD building production-scale GenAI, NLP, and Document AI systems for enterprise insurance clients — directly contributing to 7% of company ARR.

IIT Dharwad alumnus (JEE Top 1%). Ushur Champion Award recipient. Expert in LLMs, RAG architectures, transformer fine-tuning, and MLOps — with 4.5 years of hands-on experience at scale.

"I don't build models.
I architect systems that
run on them."

Case Studies

LangChain GPT-4 NER Azure

RFP Processing Pipeline

Problem: Insurance RFPs took 5 days to process manually

Outcome: 12 minutes. P95 SLA. 99%+ efficiency gain.

Azure Form Recognizer GPT-4 Pinecone

Multi-Agent Document Intelligence

Problem: 4 enterprise clients, 6 document types, no consistency

Outcome: 90–94% extraction accuracy. 100K+ docs/month.

LLM-as-a-Judge Embedding Similarity HITL

Confidence Estimation Engine

Problem: No way to know which LLM extractions to trust

Outcome: Auto-approve queues. 30% increase in daily platform usage.

Core Competencies

LLMs & GenAI

GPT-3/4, LangChain, LlamaIndex, DSPy, LiteLLM, OpenAI/Anthropic API, Microsoft AI Foundry

Models & Architectures

Llama 2/3, BERT, RoBERTa, mBERT, SpanBERT, BART, ULMFiT, FastText, YOLOv3

Fine-tuning & Techniques

LoRA, QLoRA, PEFT, RAG, Prompt Engineering, Few-shot Learning, Transfer Learning, Ontology-driven Extraction

MLOps & Infra

Docker, Kubernetes, AWS (SageMaker, EC2, S3, Lambda), Azure OpenAI, MLflow, W&B, GitHub Actions, Grafana, ONNX Runtime

Vector & Databases

Pinecone, Weaviate, ChromaDB, FAISS, Elasticsearch, PostgreSQL, MongoDB

Programming & Data

Python (NumPy, Pandas, scikit-learn), PyTorch, TensorFlow, HuggingFace Transformers, Apache Spark, FastAPI, SQL, R

Writing

Building a confidence layer for LLM extractions in production

How we combined LLM-as-a-Judge with embedding similarity to separate auto-approve from human-review

Why we replaced TF-IDF with ULMFiT for email classification

The performance gap was real. Here's what changed and what didn't.

INT8 + ONNX: cutting inference latency 50% without accuracy loss

350ms → 175ms. The quantization path that actually worked in production.

Available for senior IC roles, consulting, and technical advisory.

[email protected] +91-7013246322 linkedin.com/in/kushal-kumar github.com/kushal1999seemakurthi Bangalore, India

Download Resume ↓