Alankrit Gupta

I build machine learning systems that scale; from robotics production floors to large-scale generative AI platforms.

My work sits at the intersection of modeling, infrastructure, and reliability, with a focus on systems that don’t just train models, but continuously improve them in real-world environments.

Currently at Suki:

- Re-architected training infrastructure to achieve 97% faster cycle times (2 hrs → 3.5 min) while reducing idle storage costs by 85%.

- Led GPT-4 → Gemini 2.5 Pro migration analysis, modeling token consumption and GSU capacity to ensure scalable throughput.

• Developed distributed RL training pipelines and custom reward functions for automated clinical-note evaluation.

Previously at Dexterity:

- Designed a production Prediction Monitoring system for instance segmentation models that improved mAP by 12% and reduced false negatives by 5–7% through targeted retraining workflows.

- Standardized debugging frameworks for 24/7 robotics fleets, reducing average remote recovery time to <15 minutes.

Technically, I work across distributed training (DeepSpeed, Accelerate, Unsloth), RL-based evaluation design, and cloud-native ML systems on GCP and AWS. I’m particularly interested in ML systems engineering, large-scale model training, and infrastructure that balances performance, cost, and reliability.