Tanish Mishra

B.S. Computer Science, Minor in Data Science

Revmo AI

• Engineered a custom time-series analytics engine supporting user-defined computed metrics, utilizing Django ORM annotations for JSON extraction and topological sort algorithms to safely resolve metric dependencies.
• Diagnosed systematic failure modes in voice-agent dialog by analyzing 1,000+ call transcripts; rewrote core system prompts to cut average order completion latency 25% (∼ 60s) and reduce monthly LLM inference spend.
• Guided production model selection for national clients by designing evaluation suites and benchmarking 19 LLMs and 4 STT providers (Deepgram, Azure, etc.) across transcript accuracy, turn segmentation, and latency.
• Fine-tuned a 1B-parameter Llama with PEFT/LoRA to generate context-aware filler phrases masking LLM inference latency in real-time voice calls, replacing 8B model with no measurable degradation in dialog quality.

ARC Lab, Arizona State University

• Improved multi-GPU throughput for adversarial Teacher-Student RL experiments by engineering a custom Python L-Uniqueness Creativity Index and integrating live rewards into a distributed Ray pipeline.
• Diagnosed reward gaming and tokenizer misalignment on Phi-4 by conducting systematic Weights & Biases ablation studies, tracking performance across 20+ automated benchmarks including AIME and GPQA-D.

Mintelium

• Developed a React Compliance Management (CaaS) dashboard for easy compliance oversight by consolidating key metrics, automated workflows, and progress tracking into one view.
• Automated GDPR workflows via AWS Textract/Macie, replacing manual compliance review.
• Engineered AWS data pipeline (Step Functions, Lambda, S3), cutting processing time by more than 50%.
• Enhanced application usability by identifying and implementing targeted front-end UX/UI improvements.

Reddit cross community analysis topic modeling pipeline including crawling, BERTopic analysis, and LLM-assisted visualization.

AI Assistant tuned on base llama 3.2 3B model for instruction following and safety alignment.

A Hybrid CNN plus Transformer Image Captioning model

Spotify song analysis: Examining musical attributes vs. streaming success and trends from 1930-2023. R used for statistical modeling and visualization. Key finding: platform presence matters more than song characteristics.