Paras Jain
DevOps Engineer
@ medecro.ai
BUILT
- Started Talos
- App and database management
- Ran medecro.ai's infrastructure on it
- Product direction
THE TEAM · TALOSFLEET
Three people built Talos.
Talos started as one engineer’s answer to too many tools at medecro.ai, and grew into the control plane that runs the company’s infrastructure. These are the three people who built it, and the part each of them built. Talos belongs to TalosFleet, the three of us; medecro.ai is where it was born and still runs.
THE THREE OF US
Aryan Mishra
Backend & System Architect
@ medecro.ai
BUILT
- The layered architecture for concurrency, caching and availability
- Role-based access control that holds across an organisation
- The job processor every action runs through
- Database tuning and security
- The MLOps manager
Vaibhav Rana
AI Engineer
@ medecro.ai
BUILT
- Agent and MCP management
- The MLOps pipeline design
- The agent behind the AI DevOps Engine
HOW TALOS WAS BUILT
Five chapters.
- 1
Too many tools
Paras, then a backend engineer at medecro.ai, was asked to bring DevOps to the team. The work was scattered across a dozen tools, and every routine task meant visiting several of them. So he started a small project to manage apps and databases in one place, and it began running the company's infrastructure.
Paras - 2
Built for scale
When the company moved to Docker and Kubernetes, with production on AWS and dev on Azure, the project grew with it. Aryan designed the layered architecture that lets it handle a steady stream of deployments concurrently: caching, availability, a job processor every action runs through, optimised databases and the security around them, and role-based access control that still holds when a whole organisation shares one install.
Aryan - 3
Running a company
At that point two people's tool ran medecro.ai's infrastructure: two Kubernetes clusters, five databases, nine served AI models, the company's own VPN, three client applications and two admin applications.
ParasAryan - 4
Agents, MCP and models
The company's AI agents connected tools like Notion and GitHub, so Talos gained agent and MCP management. Vaibhav joined with hands-on knowledge of how agents and MCP servers run in production. With medecro.ai's nine models trained on Colab and tracked in Supervisely, he proposed an MLOps manager and shaped its pipelines; Aryan designed and built it.
VaibhavAryan - 5
The AI DevOps Engine
Watching what Claude Code made possible, Vaibhav suggested an AI engine for DevOps. Training our own model would have needed far more data and GPU budget than we had, so we built an agent on top of existing AI providers instead: a retrieval layer gives it the right context, and a telemetry layer learns from the decisions people make. Vaibhav designed the agent.
Vaibhav
IN PRODUCTION AT MEDECRO.AI
A typical week.
medecro.ai is Talos’s first user and still runs on it. Every figure below is per week.
15–20deployments a week
150+users a week
20k+logs collected a week
10k+requests managed a week
32k+database queries a week
WHAT IT RUNS
2Kubernetes clusters
5databases
9AI models served
5applications (3 client, 2 admin)
1company VPN
2clouds (AWS prod, Azure dev)
Free to self-host, on a machine you already have.
Install Talos and point it at one server. The three of us read every message.