Harshit Garg

AI mock interview platform

Hirely

• Built a full-stack AI mock interview platform supporting resume screening, ATS evaluation, and real-time voice interviews using Django, REST APIs, WebSockets, Deepgram, Groq, and ElevenLabs.
• Designed a low-latency speech-to-speech pipeline enabling conversational AI with streaming transcription, LLM inference, and neural text-to-speech.
• Implemented Retrieval-Augmented Generation (RAG) using MiniLM sentence embeddings for resume-aware question generation and contextual skill assessment.

Controlled NLP experiment

Context Window Length Analysis using GPT-2

perplexity

−91%

perplexity

answer similarity

+78%

answer similarity

latency

9ms

latency

• Fine-tuned GPT-2 on the SQuAD dataset across four context window sizes (32–256), keeping other variables constant to isolate the impact of context length on QA performance.
• Achieved a 91% reduction in perplexity and 78% improvement in answer similarity (cosine similarity) when scaling context length from 32 to 256 tokens.
• Measured 9ms latency across all configurations and identified 256 tokens as the optimal balance between answer quality and computational efficiency.

Classical ML + explainability

Network Intrusion Detection System

macro F1

+16%

macro F1

R2L recall

13×

R2L recall

features

78

features

• Classified 5 network attack categories on the NSL-KDD dataset across 78 traffic features, benchmarking XGBoost, RandomForest, and KNN, with XGBoost achieving the best overall performance.
• Applied SMOTE to handle class imbalance, boosting macro F1 by 16% and R2L recall by 13×.
• Analyzed feature importance using SHAP, uncovering count and dst_host_srv_count as dominant signals for DoS and normal traffic classification respectively.