2026Researchresearch
Emotion Engine
A 72-feature LSTM rediscovers fear, grief, and suspicion from zero hardcoded rules.
p = 3.3e-113 across 205,940 agent-step records
PyTorch · Python · Django · Custom sim engine
Currently in New YorkOpen to relocation
Prem Babu Kanaparthi
AI/ML Engineer building reliable, scalable AI systems across LLMs, inference, agents, and applied ML.
Currently building: reliable AI agents, inference systems, and evaluation tools for production LLM workflows.
Selected systems I've built across inference, agents, and ML evaluation.
2024Productionshipped
Multi-Model LLM Routing
Production routing across 3+ foundation models at 1.5s end-to-end latency.
18% inference cost ↓ · 42% incidents ↓ · 35% MTTD ↑
LiteLLM · AWS Bedrock · SageMaker · CloudWatch
2025Toolinglive
Sushi
Premium EDA platform — automated data quality, stats, and visualizations.
95% faster analysis · A–F quality grading · 6+ formats
FastAPI · Next.js 14 · Plotly · Clerk
Where I've shipped.
Feb 2024 — Jul 2024
Newark, CA
Generative AI Engineer @ Concentrix + Webhelp
Built a production multi-model LLM routing layer behind enterprise apps, owning latency, cost, and reliability.
- Designed a routing system across 3+ foundation models with 1.5s end-to-end latency and 18% inference cost reduction.
- Integrated LiteLLM with AWS Bedrock and SageMaker, handling thousands of daily requests at 95%+ response accuracy.
- Shipped CloudWatch-based observability with logging, safety checks, and drift monitoring — cut prod incidents 42% and improved MTTD 35%.
Aug 2023 — Jan 2024
India
Data Science Intern @ AlphaBits Technologies
Rebuilt the ML experimentation and evaluation pipeline for ranking models.
- Redesigned preprocessing and evaluation across 5+ model variants, cutting iteration time 90% and lifting ranking relevance 10%.
- Stood up a centralized feature store and reusable training pipelines for 100% reproducible experiments.
Jun 2023 — Aug 2023
India
ML Engineer Intern @ iNeuron.ai
Built supervised classifiers for phishing detection and productionized them.
- 20+ engineered URL/domain features; 92% classification accuracy on held-out validation.
- Wrapped models in production-style services with logging, monitoring, and validation checks.
Published research.
2024
Preprint
Lightweight Channel Attention for Efficient CNNs
Prem Babu Kanaparthi
Designed and evaluated a lightweight channel attention module (LCA) achieving competitive accuracy with negligible parameter and latency overhead on ResNet-18 and MobileNetV2.
Building something interesting?
Open to AI and ML Engineer roles, research collaborations, and the occasional weird side project. The fastest way to reach me is email.