mohammad-ummair — Curriculum Vitae
Mohammad Ummair

Mohammad Ummair

ML Systems & Distributed Training/Inference Infrastructure

mohdummair.iitb@gmail.com +91-8446856277 LinkedIn GitHub Blog Google Scholar
Curriculum Vitae

Research Interests

I work on ML systems for large-scale distributed training and inference. Most of my interest sits in a few places: model parallelism (both intra- and inter-operator), cutting the communication cost of tensor computation across devices, and the serving side, where throughput and scheduling decide whether an LLM is actually usable at scale. I've also been thinking about load balancing for long-context training, which tends to break the assumptions that simpler schemes rely on. Alpa, vLLM, and DistServe are the systems I keep coming back to. What I want to build is infrastructure that makes training and deploying large models cheaper without giving up scale.

Education

Indian Institute of Technology (IIT), Bombay Jul 2021 – Jul 2023
Master of Technology, Computer Science & Engineering · CPI: 8.74/10
Mumbai, India
  • Thesis: Enhancing Robustness of Smart Contracts Through Declarations (Advisor: Prof. R.K. Shyamasundar)
  • Admission: GATE CS 2021 — All India Rank 95 out of 101,922 candidates (Top 0.1%)
  • Relevant Coursework: Advanced Blockchain Technology, Foundations of Machine Learning, Algorithms & Complexity, Design & Engineering of Computing Systems, Introduction to Blockchains & Cryptocurrencies
Government College of Engineering, Aurangabad Jul 2017 – Jul 2021
Bachelor of Engineering, Information Technology · CPI: 7.5/10
Aurangabad, India
  • Achievement: Winner, Smart India Hackathon 2020 (Process Modelling, GAIL India Pvt. Ltd.)

Publications

Enhancing Robustness of Smart Contracts Through Declarations
R.K. Shyamasundar, Mohammad Ummair, S. Borse
Springer, 2024
  • Declarative Security Framework: Proposed a novel Solidity language extension (Solidity+) that abstracts run-time monitoring via high-level declarations (NONRENTRANT, PARALLEL, ACCESS, INVAR), enabling automatic instrumentation of security logic at compile time without altering programmer-visible contract structure.
  • Transformation Pipeline: Designed and implemented a source-to-source preprocessor that semantics-preservingly transforms Solidity+ programs into hardened Solidity, embedding assertions as an on-chain proof skeleton — realizing a proof-carrying code paradigm for immutable blockchain contracts.
  • Vulnerability Taxonomy & Algorithmic Mitigations: Formally characterized six Solidity vulnerability classes — reentrancy, arithmetic overflow/underflow, unchecked send, tx.origin misuse, block.timestamp dependence, and concurrency races — and developed targeted transformation algorithms for each, covering critical failure modes responsible for the $60M DAO hack and $30M Parity wallet exploit.
  • Concurrency & Access-Order Specification: Introduced PARALLEL and NONRENTRANT declaration clauses to formally specify and enforce concurrency constraints at the language-semantics level, enabling coarse-grained concurrent invocation policies and non-reentrant guarantees without manual mutex instrumentation.
  • Comparative Evaluation: Empirically evaluated Solidity+ against leading static analysis tools (Mythril, Oyente, Manticore, Porosity) on real-world smart contract datasets, demonstrating that a declaration-driven run-time enforcement approach addresses vulnerability classes that purely static methods cannot soundly detect.
View Publication

Writing

I keep a blog for working notes — ML systems, inference, and silicon. Mostly things I've measured, read, or argued about.

AI Slop Aug 16, 2026

Coding assistants feel effortless. The evidence says they make us write more code and work slower, and the dependency that builds only benefits one party.

aisoftware engineeringopinion
All posts

Industry Research & Engineering Experience

Apple Inc. Jul 2023 – Present
Software Engineer — Systems & Infrastructure
Bangalore, India
  • Distributed Task Execution Framework: Designed and implemented a multi-node parallel execution system for a firmware validation pipeline serving 100+ engineers. Engineered task scheduling, inter-node communication, and load distribution logic, achieving a 30% reduction in end-to-end latency — directly analogous to worker-level scheduling in distributed ML training systems.
  • LLM Inference Pipeline Engineering: Designed and deployed an end-to-end LLM inference serving system on Apple Foundation models, engineering in-memory batching, query routing, and retrieval-augmented generation pipelines achieving 30% higher accuracy and 40% faster query resolution.
  • Queue-Based Scheduling & Latency Optimization: Built a message-queue-driven task dispatch system (RabbitMQ) with incremental execution logic that reduced average task wait time from 100s to 10s by eliminating redundant work — structurally similar to straggler mitigation in pipeline-parallel training.
  • MCP Server for Distributed Framework Access: Designed and built a Model Context Protocol (MCP) server exposing internal distributed execution framework APIs to LLM-based agents, enabling engineers to interact with and orchestrate distributed tasks through natural language.
  • Infrastructure Observability: Developed real-time telemetry and structured logging in Go and Python across distributed node infrastructure, sustaining 99.9% system reliability in production.
Tata Consultancy Services Feb 2021 – May 2021
Project Intern
India
  • Time Series Forecasting: Deployed hierarchical time series models to forecast client export demands, enabling proactive inventory management and improving demand accuracy for daily operations.
TechInvento May 2019 – Jun 2019
Machine Learning Intern
India
  • Computer Vision for Document Digitization: Built an automated CV pipeline to digitize handwritten documents, saving the client an estimated $1.5M/month; deployed as a Node.js web application for production monitoring.

Research Projects

High-Throughput LSM-Tree Storage Engine Code

Engineered a persistent Log-Structured Merge-tree optimized for write-heavy workloads with Bloom filter-based I/O elimination and tiered compaction — building intuition for storage-compute trade-offs central to ML checkpoint and activation management.

Single-Machine Linearizable Key-Value Store Code

Built a fault-tolerant KV store guaranteeing linearizable consistency and exactly-once semantics under failures via write-ahead logging and operation identifiers — grounding understanding of consistency models relevant to distributed parameter servers.

Distributed Ledger & Consensus Simulation

Simulated a multi-node P2P network with Proof-of-Work consensus and transaction propagation; benchmarked throughput, latency, and resilience under adversarial conditions, developing hands-on experience with distributed coordination protocols.

Multi-Threaded KV Server over gRPC (Course: CS744)

Implemented a concurrent client-server KV store with pluggable caching strategies and a load generator to study throughput/latency under increasing concurrency — mirroring performance analysis methodology used in ML systems evaluation.

Asset Performance Monitoring with ML (Smart India Hackathon 2020 — Winner)

Developed a real-time APM application integrating ML to optimize mechanical device performance for industrial use; selected as winning solution for GAIL India Pvt. Ltd.'s process modelling problem statement.

Teaching & Academic Service

Teaching Assistant — IIT Bombay, Department of Computer Science
Mumbai, India · 2021–2023
  • CS 765: Introduction to Blockchains, Cryptocurrencies & Smart Contracts (Prof. Vinay Ribeiro) — Conducted lab sessions, graded assignments and exams, and provided mentorship to graduate students on blockchain concepts and smart contract development.
  • CS 744: Design & Engineering of Computing Systems (Prof. Mythili Vutukuru) — Administered and graded lab assignments, conducted vivas (oral examinations), and mentored students on systems design principles.
  • CS 251: Software Systems (Prof. Amitabha Sanyal) — Assisted with lab sessions, grading, and student mentorship.
  • CS 254: Digital Logic Design Lab (Prof. Virendra Singh) — Supported practical lab instruction and student evaluations.
Prime Educator — GATE CS Preparation
Unacademy · May 2021–Jun 2021
  • Mentored a cohort of 20 students for GATE CS; designed customized daily practice problems and performance improvement plans for underperforming students.
Department Placement Coordinator & Web Secretary
IIT Bombay, Department of Computer Science · Jul 2022–Jul 2023
  • Managed end-to-end campus placement logistics for the Computer Science department across 100+ recruiting companies.
  • Maintained and updated official Computer Science department web presence.

Technical Skills

Languages
PythonC++Go JavaScript/TypeScriptSQL JavaC
Frameworks
FastAPIFlaskReact Node.jsHaystackLangChain Hugging FacegRPC
Infrastructure
AWS (EC2, S3)DockerKubernetes PostgreSQLMongoDBRabbitMQ CI/CDGit
Research Tools
LaTeXMATLAB SolidityEVM tooling