Who I Am — Umang Sharma

Overview

I'm Umang Sharma, a Computer Science student graduating in 2027.

I'm primarily interested in AI systems and developer infrastructure — especially what happens beyond simply calling an LLM API.

My interests sit at the intersection of:

• AI/LLM Engineering
• Agentic Systems
• ML Infrastructure
• Developer Tooling
• Backend Systems
• Local AI
• Web3 Infrastructure

I tend to learn by building things, breaking them, figuring out why they broke, and then rebuilding them.

---

Current Direction

My strongest current interest is local and agentic AI systems.

I've been experimenting with running increasingly capable models locally on Apple Silicon, including:

• Qwen
• Qwen-Coder
• MLX
• Ollama
• ONNX Runtime
• Quantized models
• Model serving
• MCP
• Agent tools and skills
• Model routing
• Local coding assistants

I currently use an M2 Max MacBook Pro with 32 GB RAM as my primary development machine and have been pushing it fairly hard with local AI workloads.

I've experimented with models ranging from small coding models to 30B+ parameter models, including quantized inference and different serving approaches.

This has given me practical exposure to:

• VRAM / unified-memory constraints
• Quantization
• Model loading
• Inference latency
• Context size
• Metal memory limitations
• Model serving
• Process management
• Local model routing

---

AI & LLM Engineering

I've worked across several layers of the AI stack.

Models

I've experimented with:

• Qwen
• Qwen-Coder
• QLoRA
• PEFT
• Hugging Face
• MLX
• Ollama
• bitsandbytes
• Transformers

I've also experimented with fine-tuning smaller models and running quantized models locally.

RAG

I've built around:

• Embeddings
• Gemini Embedding 001
• Qdrant
• LangChain
• Retrieval pipelines
• Vector search

Agents

I've explored:

• Multi-agent architectures
• CrewAI
• Agent orchestration
• Tool calling
• MCP
• Agent skills
• Model routing
• Local agent execution

One of my recurring interests is:

How do you build an agent that is actually useful rather than simply wrapping an LLM in a chat interface?

---

Aegiron

One of the projects closest to what I want to work on is Aegiron, an AI-native desktop assistant.

The concept is a local-first assistant that can operate across the desktop rather than existing purely as a chat window.

The architecture I've explored includes:

                 ┌───────────────┐
                 │     User      │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │    Aegiron    │
                 │     Agent     │
                 └───────┬───────┘
                         │
             ┌───────────┼───────────┐
             ▼           ▼           ▼
          LLMs         Tools        Skills
             │           │           │
             ▼           ▼           ▼
          Ollama        MCP       Local Apps
             │
             ▼
        Local Models