Skip to main content

Selected work

Agentic systems, training runs, and research.

AI Gateway
Abstract cover for the PurpleMerit AI gateway

A gateway that routes each prompt to the model tier the task actually needs.

Department metadata, context-window heuristics, and an in-house complexity classifier cut average inference cost by 27%.

AI Engineer, PurpleMerit, 2026

ZysecAI
Abstract cover for the ZysecAI inference work

A 1.5B coder model that replaced a 70B call for production JSON.

Fine-tuning Qwen 2.5 Coder 1.5B on cached telemetry cut inference latency by 11x and token cost by 96%.

SDE-1 AI, ZysecAI, 2025

Story Teller
Abstract cover for the GPT story teller

A 124M GPT trained from scratch until it could write multi-sentence stories.

Trained on TinyStories with FlashAttention, bf16, and tensor parallelism. Throughput rose from 14k to 41k tokens/s, and validation loss landed at 1.1.

Independent, 2025

Denoiser
Abstract cover for the endoscopy denoiser

Diffusion denoising for endoscopy images, distilled to about 10x faster sampling.

Progressive distillation kept PSNR, SSIM, CNR, and lesion-detection sensitivity roughly in place while cutting inference time.

Research Intern, Samsung R&D, 2023

Let’s connect

Open to research, inference work, and teams building agentic systems. Reach out.

2026 © Gopichand Madala