Navin Kumar

I’m an AI/ML Engineer with 1 year of experience fine-tuning open-source transformer models and shipping end-to-end ML services. I focus on supervised fine-tuning, structured outputs, evaluation systems, and model failure analysis to make models more reliable in real products.

In my current role at DigitalxCode, I built an agentic stock intelligence platform using GPT-4o and a custom MCP orchestration layer, producing GREEN/AMBER/RED investment signals across 24 equity buckets daily with zero human intervention. I implemented a RAG memory pipeline with ChromaDB and OpenAI embeddings to reduce redundant recommendations by injecting semantic historical context before health checks, and I deployed a containerized full-stack system on GCP Compute Engine with FastAPI, Next.js, PostgreSQL, Redis, and ChromaDB.

I also fine-tune and optimize transformer models for performance-critical applications. I trained an RT-DETR detector on the VisDrone dataset and improved measurable benchmarks like mAP, while reducing inference latency to under 33ms (30+ FPS) using model optimization, batching, and runtime profiling; I paired this with an InsightFace + OpenCV face recognition pipeline for live video throughput. Previously, I owned an end-to-end time-series forecasting service deployment (data exploration, model selection, Docker packaging, GCP deployment) and integrated LLM-based workflows into downstream analytics for structured business intelligence outputs.