Yashasvi Udayan — AI Engineer

Yashasvi Udayan

AI Systems Architect

I design and ship production-grade AI systems — multi-agent intelligence architectures, LLM orchestration engines, and autonomous DevOps pipelines. Built on OpenAI, Anthropic, and open-weight models. Zero to deployed, at velocity.

View ProjectsGet in Touch

Work

Experience

Independent AI Systems Developer

Jan 2026 – Present

Self-Directed · Lucknow, India

  • Designed and shipped five production AI systems: The Orchestrator (multi-agent natural language CLI), PR Agent (automated code review via local LLMs), Context Core (persistent vector memory layer), The Autonomous Researcher (self-directed web research pipeline), and Sentinel-Shield (enterprise AI security gateway) — all built on local compute or open-weight models.

  • Built Sentinel-Shield to solve the enterprise Trust Gap: a security gateway with real-time data leak interception, gateway-level prompt injection defense, per-user budget enforcement, and permanent compliance audit trails targeting Finance and Healthcare.

  • Mastered Apple Silicon (Metal) GPU acceleration to run 8B parameter LLMs locally at $0 cloud cost; established CI/CD via GitHub Actions with multi-platform Docker image publishing and security-first credential redaction across all vector ingestion workflows.

Technical Skills

AI & Machine Learning

LangGraphRAGVector Databases (ChromaDB)Agentic OrchestrationLocal LLMs (Ollama, Llama-3-8B, Nomic)Prompt EngineeringMulti-Agent Systems

Security & Governance

AI Security GatewaysPrompt Injection DefenseData Leak PreventionCompliance Audit SystemsBudget Governance

Backend & Infrastructure

PythonFastAPIRedisDocker (Multi-platform)GitHub Actions (CI/CD)Pydantic v2

Hardware & Tools

Apple Silicon (Metal) accelerationLocal inference optimizationGitGitHub CLISQLitePytest

Work

AI Projects

Production-grade systems built with local LLMs, autonomous agents, and real-world constraints.

Swarm-Tune

In Progress

A decentralized AI fine-tuning swarm — nodes discover each other over libp2p, exchange raw PyTorch gradients across the mesh, average them without any central server, and collectively fine-tune a model inside a Docker-simulated network on a single machine.

  • Phase 1: Whisper Network — libp2p P2P node discovery & messaging, no central server
  • Phase 2: Gradient Math — manual PyTorch gradient extraction, bypassing DDP for network transit
  • Phase 3: Synchronisation — nodes exchange & average gradients across libp2p; straggler-problem handling
  • Phase 4: Docker Simulation — 4-container swarm on M4 Pro collectively fine-tuning a model
libp2pPyTorchDistributed MLDockerP2PPython

Sentinel-Shield

Live

An AI security gateway that sits between your app and OpenAI / Anthropic / Ollama — stripping PII, blocking prompt injections, scanning responses for leaks, enforcing rate limits, and emitting a full audit trail. Drop-in proxy, Docker-ready.

SecurityAI GatewayDockerPrometheusPythonSQLite
GitHub

The Orchestrator

Live

A natural language interface to control and orchestrate multiple AI agents from a single terminal. Think tmux meets AI — dispatch tasks, monitor pipelines, and collect results.

Multi-AgentCLIOllamaPython
GitHub

PR Agent

Live

Automated code review agent that analyzes pull requests, suggests improvements, detects potential bugs, and generates meaningful commit messages using local LLMs.

Local LLMGitPythonFastAPI
GitHub

Context Core

Live

A persistent memory layer for AI assistants. Stores, retrieves, and compresses conversation context using vector embeddings to give LLMs long-term memory without token bloat.

Vector DBEmbeddingsRAGTypeScript
GitHub

The Autonomous Researcher

Live

An AI agent that autonomously browses the web, synthesizes information, and produces structured research reports. Powered by Llama-3 via Ollama with a multi-step planning pipeline.

Llama-3OllamaLangChainPython
GitHub

About

Building AI systems that ship.

I'm an AI Systems Architect who designs and builds production-grade intelligent systems — from cloud-deployed LLM applications to autonomous agent pipelines and internal developer tooling.

My work spans OpenAI & Anthropic APIs, open-weight models, and multi-agent orchestration. I care about the full stack: architecture, reliability, and user-facing impact.

When I'm not shipping AI products, I'm deep in systems design, evaluating new model capabilities, or experimenting with agent reasoning frameworks.

4+
AI Projects
Full
Stack AI
Prod
Ready

Environment

Dev Stack

Hardware and tooling that power day-to-day AI development and experimentation.

Machine

Apple M4 Pro

14-core CPU · 16-core GPU · 24 GB Unified Memory

LLM Runtime

Ollama

Llama-3.1 · Mistral · Phi-4

Primary Model

Llama-3.1 70B

Quantized Q4_K_M · Local inference

Daily Tools

Claude CodeAI Coding Assistant
CursorIDE
WarpTerminal
Docker DesktopContainers
TablePlusDatabase GUI
ProxymanNetwork Inspector

Learning

Courses and programs actively shaping my technical foundation.

Contact

Let's build something.

Open to AI engineering roles, freelance projects, and interesting collaborations. I respond within 24 hours.

GitHubLinkedInX