ABOUT — WHO I AM

Engineer ofintelligent systems.

I'm Vatsa Joshi — an AI/ML engineer in Bengaluru focused on LLM evaluation, fine-tuning and agentic systems. 2+ years shipping production LLM systems, including a founding AI engineer role owning retrieval and memory infrastructure end-to-end.

I've built and owned evaluation harnesses for production LLM pipelines — including a legal NER system validated to 96% document-level zero-failure across a 4M-row eval set while cutting extraction cost 25x over an 8M-judgment corpus. Hands-on with LoRA/QLoRA, DeepSpeed ZeRO, red-teaming, and adversarial/deepfake detection (IEEE-published). High-ownership; ships to production.

CAPABILITIES
LLM Evaluation & Reliability
Eval pipeline design · Golden/regression sets · Synthetic test generation · LLM-as-judge · Failure taxonomies · Grounded extraction · Observability & tracing
Fine-Tuning & Training
LoRA · QLoRA · Unsloth · HF Transformers/PEFT/TRL · DeepSpeed ZeRO · Accelerate · Multi-GPU training · Quantization · Dataset curation
Agents & Retrieval
LangChain · LangGraph · CrewAI · FastMCP · RAG/CAG · FAISS · Qdrant · Elasticsearch · GraphRAG (Neo4j) · Tool-calling & routing
AI Security & Adversarial ML
Prompt-injection & jailbreak testing · Red-team case generation · Adversarial detection (EfficientNet-B7 + Bi-LSTM) · OWASP LLM Top 10 · MITRE ATLAS · NIST AI RMF
ML & Vision
PyTorch · TensorFlow · Keras · Scikit-learn · XGBoost · LightGBM · OpenCV · YOLO · SSD · PaddleOCR · Diffusers
Systems & Infra
Python · Rust · TypeScript · FastAPI · NestJS · Docker · AWS · GPU inference · PostgreSQL · MongoDB · Redis
EDUCATION
B.Tech, Computer Science
Navrachana University, Vadodara
AUG 2022 — MAY 2026
PUBLICATIONS & AWARDS
IEEE-published — Deepfake Detection
EfficientNet-B7 + Bi-LSTM · IEEE Xplore, DOI 10.1109/CCIC68129.2026.11486188
2nd place — Tinkerthon 3.0 + 1-year SSIP Gujarat research grant
Top 10 — Odoo x NMIT Hackathon · 2 papers on ResearchGate