vllms

A CPU Realtime VLM in 500M. Surpassed Moondream2 and SmolVLM. Training from scratch with ease.

Proxima lets existing GPUs serve 4x more concurrent requests

My complete journey to becoming an Agentic AI Engineer through structured learning, projects, experiments, and production-ready implementations of modern AI systems.

NeuralForge is an end-to-end LLM fine-tuning & serving platform. It leverages LoRA (PyTorch/Hugging Face) for domain adaptation, cutting error rates vs. base models. Features an automated MLflow eval harness tracking accuracy, latency & hallucinations with AI guardrails, deployed on vLLM, Docker & Kubernetes via CI/CD for high throughput.

WikiMixQA: A Multimodal Benchmark for Question Answering over Tables and Charts [ACL 2025]

Gemma4-12B multimodal specialist on DGX Spark: image input, voice-note workflows, tool calling, 196K tested context.

Multi-agent creative pipeline — Director, Scriptwriter, and Cinematographer agents coordinate over Veo3 on Vertex AI to generate 10 short-form promo videos from raw brand assets. Gemini orchestration + vLLM inference.

Phani's Portfolio