Ramakrishna Bachu

About Ramakrishna Bachu

I train and serve small language models end-to-end, from dataset generation through quantized inference.Shipped artifacts: NPC Fast 1.7B (SmolLM2 base, full-weight continued pretraining to 16K context, router LoRA SFT, 98.3% out-of-distribution routing accuracy). NPC Fin 32B (Qwen2.5 base, 32,496 SFT examples covering 59.7M tokens across a five-category taxonomy, distributed training across 12xH100, GPTQ 4-bit quantization to 19.2GB). NPC Fin PRM 7B (process reward model for step-level verification, Spearman 0.94, Error F1 0.87 on held-out set). All public on HuggingFace.Serving: Mixture-of-Models inference API on vLLM + FastAPI, RunPod A40, auto-routing between model sizes, MongoDB-backed auth with tier-based rate limiting. Live at bottensor.xyz.Research interests: post-training (SFT, DPO, GRPO), reasoning models, process reward models, efficient inference, small model distillation.Software Engineer at Lowe\'s building production Java/Spring Boot distributed backends and API performance infrastructure.

Experience

1. Software Engineer

   Lowe's Companies, Inc.

   Jul 2023 — Present · NC, US

   Reduced API response time from 200ms to 50ms on Spring Boot microservices serving high-traffic e-commerce backend- Built and deployed Kafka consumers for registration and promotion event pipelines- Designed performance testing suite with JMeter; established latency regression benchmarks across service releases- Deployed and monitored services on Google Cloud Platform (Apigee, Spinnaker) with Prometheus and Grafana observability.

Education

• Jawaharlal Nehru Technological University

  Bachelor’s Degree, Information Technology

  2012 — 2016

• University of New Haven

  Master of Science - MS, Computer Science