Naman Ahuja

Naman Ahuja is a software engineer at Meta working on new hardware adoption and the distributed infrastructure supporting large-scale AI workloads. His work connects hardware integration, capacity management, autoscaling, and reliability: making new compute available across a datacenter fleet while maintaining the performance and reliability that production services require.

From applied machine learning to production infrastructure

Ahuja’s earlier projects explored practical problems in applying machine learning to information-intensive tasks. In 2018, he collaborated on a system for generating chapter-level summaries of theses and dissertations. The project faced two obstacles before summarization could work: extracting well-structured chapter text from PDFs and finding enough labeled academic documents to train models. The team compared scholarly PDF extraction tools and used transfer learning from related Wikipedia articles to address the training-data shortage, experimenting with supervised and reinforcement-learning-based summarization models.

His 2019 projects included reinforcement-learning approaches to scheduling news-website crawls and adversarial testing of hate-speech classifiers. The crawling project sought to predict changes in page content and website structure. The classifier project tested whether small text perturbations could evade detection while preserving the original meaning, exposing weaknesses in deep-learning-based moderation systems.

His research also examined human–AI collaboration. The 2020 paper Response Quality in Human-Chatbot Collaborative Systems, described in his publication record, studied how people composed messages using chatbot suggestions. Two collaborative systems encouraged different degrees of synthesis between human writing and AI output. Both produced more informative responses with less user effort than a human-only baseline, while sacrificing some fluency and human likeness. Compared with chatbot-only responses, the collaborative systems retained comparable informativeness while producing more fluent, human-like messages. The results exposed a design tradeoff: assistance could improve the information in a response without improving every aspect of its quality.

At Meta, Ahuja works on the infrastructure supporting AI computation. He has led initiatives to integrate new hardware into the datacenter fleet, improve compute utilization, and enable large-scale workloads. His speaker profile describes work spanning capacity management, autoscaling, reliability engineering, and datacenter-scale resource scheduling. It also includes infrastructure and tooling for agentic workflows, safety guardrails for autonomous decision-making, and evaluation frameworks for real-world deployment.

Alongside Nishant Gupta, his university research collaborator and later fellow engineer at Meta, Ahuja launched BuzzingTech, a publication and mentorship platform focused on distributed systems, AI infrastructure, and technical leadership. Their launch announcement connected practical engineering knowledge with career development: understanding the systems behind AI applications and helping senior engineers take on broader technical responsibility.

Operating inference as a coordinated system

In his AI Engineer World’s Fair presentation, Operating Distributed Inference Systems at Scale, Ahuja argued that agent workflows change what inference infrastructure must coordinate. A single user request can trigger reasoning steps, retrieval, tool calls, retries, and additional model invocations. His account of the talk explains why routing, scheduling, caching, queue management, and fault isolation can determine performance before the model itself becomes the bottleneck.

• Schedule for the workflow. Ahuja argues that scheduling should account for model locality, context size, available memory, cache state, priority, and the workflow’s remaining latency budget. These factors connect resource placement to the user’s task: a decision that improves accelerator utilization can still worsen tail latency or leave too little time for subsequent steps.

• Use control loops to manage competing goals. His writing on inference orchestration describes a control plane that observes workload conditions, applies policies, and adjusts routing, admission, batching, and capacity across models and heterogeneous accelerators. The decisions interact. Reducing latency can reduce batching efficiency; increasing utilization can create tail-latency problems; overprovisioning improves responsiveness at a higher infrastructure cost. He favors continuous adjustment over static provisioning and manual tuning.

• Make observability inform operations. In his distributed-inference talk, time to first token, time per output token, batch fill rate, KV-cache hit rate, queue depth, and cost per successful task are inputs to operating the platform. Together, they help engineers assess how changes in routing, batching, and capacity affect latency, throughput, reliability, and cost.

Measuring the effort AI assistance creates

Ahuja has returned to his human–chatbot research in writing about AI-generated suggestions. Generating several drafts can reduce writing effort while adding reading, comparison, editing, and decisions. He argues that AI writing tools should measure total task time, reviewing effort, cognitive load, trust, user ownership, and preservation of personal voice alongside output quality.

That position gives the earlier research a practical application: evaluate how quickly a person reaches a response they trust and feel comfortable sending. One relevant suggestion at the right moment may help more than repeated generations, particularly when additional alternatives make the decision harder.