Mohammad Ausaf Logo Image
Mohammad Ausaf

mohammad ausaf

Agent Systems · AI Infrastructure
@ Galleri5 (Collective Artists Network)

get to know me

I build agent systems and the infrastructure they run on. I joined Galleri5 (part of Collective Artists Network) when the tech team was about 10 people, and in my first 3 years I've worked the whole depth of the AI stack, from model training to distributed systems and production ops. Today I'm a hands-on manager and one of five engineers on core AI Studio.

Most of what I do now is MIRO, our creative agent: give it a brief and it plans, generates and edits image, video and audio work, and checks what it rendered before it answers. Before MIRO, I built the canvas planner. I'm as much an infra person as a dev, so I built what runs under it too, from an empty repo: the inference gateway, fair scheduling and the code sandbox. AI Studio runs on Azure, in partnership with Microsoft.

Before agents, I worked on the models themselves: training and fine-tuning, LoRAs for character consistency, ComfyUI and PyTorch pipelines, CUDA where it mattered. Those pipelines went into Mahabharat: Ek Dharmayudh on JioHotstar and Chiranjeevi Hanuman: The Eternal (24M+ first-look views).

I also build open-source tools (cfgit, Searchy, Boring Notch Lyrics) and write about the systems I build and the philosophy of AI. If you're building agents or AI infrastructure, email me.

Email Me

Tech Stack

Python
FastAPI
PyTorch
MongoDB
Redis
Docker
GCP
Azure
AWS
Firebase
Pinecone
ComfyUI
CLIP
Google Gemini
OpenAI
Replicate
FAL AI
Agentic AI
Generative AI
Distributed Systems
MCP
Hermes Agent
OpenTelemetry
Daytona
Linux

experience

manager - software engineering

Galleri5 - Bengaluru

May 2026 - Present

Software Engineer (AI), Dec 2023 - May 2026

  • Technical lead: Co-own backend architecture and code review across AI Studio; led a team of up to 3 engineers; partner on product direction
  • Creative agent (MIRO): 50+ tools and 300K+ tool calls a month; long turns people can redirect mid-task on any replica, and the agent never authorises spend by itself. Turn completion went from about 95% to 99%
  • Evals and sandboxing: Every turn traced, with LLM judges scoring outcomes; halved content-safety false positives on a 600+ case suite that gates every push. Built the sandbox agents run code in: 80K+ runs a month, egress blocked and failing closed
  • Inference gateway: Built from an empty repo: 190K+ jobs a month to 170+ model endpoints, with deficit weighted round robin scheduling, TTL leases and crash-safe intake (it grew out of a GPU dispatcher that lost 0 of 10,000 jobs under load)
  • Billing and reliability: Usage metering that matches real cost; never re-dispatches a job that may still be running; deploys drain in-flight agent turns instead of killing them
  • Earlier: The Mahabharat episode pipeline; video embeddings and ingest hardening with Microsoft's engineering team (ISE); CLIP semantic search for Myntra, Ajio and H&M

Tech Stack: Python, FastAPI, asyncio, Redis (Lua), MongoDB, PostgreSQL, Azure, GCP, AWS, Docker, Daytona, MCP, Hermes Agent, OpenTelemetry, LangWatch, Claude, Gemini, OpenAI, PyTorch, ComfyUI

Read the full resume

open-source systems tooling

cfgit - version control for live database records

2026

  • Record history: Built a Python tool for status, diff, log, adopt, tag, and restore on records that stay in MongoDB or Postgres
  • Drift checks: Kept hashing, diffing, rollback, and stale-write detection in a DB-neutral core with datastore-specific adapters
  • Shared actions: Exposed the same operations through CLI JSON, localhost UI, MCP tools, and portable agent instructions

Tech Stack: Python, MongoDB, Postgres, MCP, CLI tooling, Localhost UI

bachelor of technology - computer science

KIET Group of Institution, Ghaziabad

2024

Relevant Coursework:

  • Probability and Statistics, Calculus, Operating Systems
  • Data Structures and Algorithms, Machine Learning
  • Databases

work

G5 AI Studio screenshot

miro, the creative agent in ai studio

MIRO is the creative agent I build at Galleri5. A brief goes in; it plans, generates and edits image, video and audio work across 50+ tools, builds review boards and production trackers inside the chat, and checks what it rendered before it answers. People can redirect a long turn mid-task, and it never authorises spend by itself. It runs on the inference gateway and code sandbox I built underneath it. Technologies: Python, FastAPI, MongoDB, Redis, Hermes Agent, MCP, OpenTelemetry.

Read More
cfgit drift review UI

cfgit - version control for live database records

Git, but for the live config rows in your database. Version, diff, drift-detect, and restore records that stay in MongoDB or Postgres, without exporting them to files. History, tags, a localhost UI, and MCP tools so an agent can safely edit config too. Technologies: Python, MongoDB, Postgres, MCP, CLI tooling, Localhost UI.

Read More
Searchy Screenshot

searchy - local ai photo search for mac

Search your Mac photos by describing them, fully on-device, nothing leaves your machine. SigLIP/CLIP semantic search with OCR hybrid ranking, so a screenshot of "invoice 2024" ranks for that query. Face grouping, duplicate detection, external volumes, and GPU-accelerated indexing on Metal/MPS. Shipped as a signed Mac app. Technologies: Swift, SwiftUI, Python, SigLIP/CLIP, OCR, Metal/MPS.

View Project
Boring Notch Lyrics

boring.notch lyrics - macOS notch music widget

Fork of Boring Notch adding real-time synchronized lyrics display. Features multiple display modes (flowing, alternating, stacked), per-monitor configuration, and timing offset controls. Supports Spotify, Apple Music, and YouTube Music. Technologies: Swift, SwiftUI, macOS APIs, Music Integration.

Read More
Deep Learning Plant Detection

deep learning based medicinal plants detection

This project aimed to identify medicinal herbs using a machine learning model(s). The model used ResNet, a type of CNN, with a validation accuracy of 98%.

Read More