Harshit Upadhyay

AI/NLP Engineer with hands-on expertise in building end-to-end LLM pipelines, fine-tuning models, information retrieval optimization, and deploying scalable agentic systems.

Professional Experience

AI NLP Engineer November 2024 - Present

Miko AI, Mumbai, India (Remote)

Junior AI NLP Engineer June 2024 - October 2024

Miko AI, Mumbai, India (Remote)

• Architected a production memory module with under-150ms hybrid retrieval (BM25 x dense embeddings via self-hosted Triton service), deployed on GCP + Kubernetes - enabling persistent user context and 3x better conversation continuity across sessions
• Fine-tuned Qwen3-4B with LoRA on 7 specialized skill domains; led a 3-pronged synthetic data pipeline (normal, direct &

indirect profanity variants) via GPT-4.1-mini, trained on a Blackwell GPU cluster with Accelerate + GRPO

• Evaluated with cosine similarity, BERTScore, tagging and task-following metrics; reduced inference latency by 70% and cut $30K+

monthly costs by optimizing serving configs, prompt efficiency, and fine-tuning pipeline at scale

• Spearheaded an autonomous red-teaming agent covering 17 safety topics (suicidal ideation, grooming, trafficking, self-harm,

etc.) - concurrent topic agents generate, attack, and score via LLM-as-judge; a context memory reflector analyses batch outcomes and sharpens the next round’s attack strategy; full overnight sweeps for $10

• Engineered a 10M+ Q&A semantic cache from a 7.6GB Wikipedia hierarchy with multi-threaded pipelines, cutting search latency by 40% and LLM inference costs by over 95%
• Established a YAML-configurable RAG/QA evaluation framework with LLM-as-judge scoring across single-turn and multi-turn conversation modes - boosting answer relevance by 30%
• Designed TIA, a multi-modal in-cabin AI assistant demoed at Tata Motors Tech Day - multi-voice TTS (Google Cloud Chirp3-HD

+ ElevenLabs), state machine-driven orchestration, and WebSocket + coturn middleware for WebRTC relay

• Applied distillation via decision tree for training data generation, boosting data quality by 60% on the animal guessing skill

Data Science Intern January 2024 - March 2024

Timely AI, Bangalore, India (Remote)

• Integrated Facebook SDK (OAuth flow, social login) into existing backend without disrupting current auth architecture
• Delivered an API-driven internal sales platform automating repetitive data operations and cutting manual reporting overhead

Machine Learning Intern May 2023 - July 2023

Canary Mail, Rajkot, India

Awarded a Letter of Recommendation from the CEO for exceptional performance

• Crafted a conversational chatbot for email interactions (summaries, reply drafts, action items), boosting user engagement by 75%
• Implemented a hybrid semantic + lexical retrieval system with reranking over email data, achieving 94% retrieval precision

Education

Indian Institute of Technology, Bombay 2020 - 2024

Bachelor of Technology, Metallurgical Engineering and Material Science

Relevant Coursework: Introduction to Machine Learning, Data Analysis and Interpretation, Computer Programming and Utilization,

Operating Systems, Computation Lab, Introduction to Numerical Analysis The New Tulip International School, Ahmedabad 2020

CBSE 12th - 96.6%

Kendriya Vidyalaya, Rajkot 2018

CBSE 10th - 94.4%