About Me

Hey Hi!, I am Mani. Nice meeting you!

I am a Software Engineer, ML at Meta (Menlo Park, CA), building and optimizing ML systems: feature pipelines, multi-agent LLM systems, and query infrastructure at scale.

Before Meta, I was an ML Research Engineer at Evidium (healthcare AI, San Francisco), where I built clinical NEL pipelines, RAG-powered chatbots, and scaled distributed training across multi-node A800 GPUs.

I hold an MS in CS (4.0 GPA) from UC Irvine, where I researched adversarial ML security on code LLMs and cognitive behavior of language models under Prof. Sameer Singh and Prof. Richard Futrell. My work was published at EMNLP Findings 2023 and AAAI EAAI 2023. I completed my bachelor's in Electronics Engineering from IIT BHU Varanasi.

I care about ML systems that work at scale and writing clearly about what I learn along the way. Outside work, I like to hike, play chess, and create art.

"You were born an original, don't die a copy." - John Mason.

Work experience

Software Engineer, ML @ Meta [Jun 2025 - Present]

  • Optimized ML feature pipelines by eliminating distributed query bottlenecks and redesigning feature materialization, saving $1.3M ARR and reducing compute by ~50%.
  • Built a Reflexion-style multi-agent LLM system where a routing agent orchestrates tool-equipped sub-agents for autonomous alert diagnosis with a self-correction loop, reducing human triage time by 80%.
  • Designed metadata-driven O(1) query optimizations, cutting latency by ~1000x on large tables and eliminating debugging bottlenecks across hundreds of daily Meta-scale queries.

ML Research Engineer @ Evidium [Aug 2023 - May 2025]

  • Architected and deployed end-to-end ML across research, training, and production for three clinical AI products at an early-stage healthcare startup.
  • Built a NEL pipeline mapping clinical text to 400K+ medical entities with constrained decoding via restriction tries.
  • Trained GNNs on a proprietary knowledge graph, fusing graph embeddings with a T5 encoder-decoder via cross-modal retrieval for entity linking; F1 improved by 10%.
  • Built a clinical AI chatbot powered by RAG with HyDE query expansion, FAISS vector retrieval, and an LLM self-critique layer for hallucination reduction.
  • Scaled distributed training across multi-node A800 GPUs using FSDP; cut inference latency by 31% via token caching and continuous batching.

Graduate Student Researcher @ UCI [Mar 2022 - Sep 2023]

  • Advised by Prof. Sameer Singh (NLP Lab, DARPA-funded) and Prof. Richard Futrell (Linguistics, NSF-funded).
  • Researched adversarial ML security on code LLMs (Incoder, PolyCoder); evaluated backdoor trigger injection and clean-label poisoning attacks under the DARPA TADS program.
  • Analyzed decision-making sensitivity of LLMs to prompt variations on the Horizon task; demonstrated human-like exploration-exploitation tradeoffs. Published at EMNLP Findings 2023.
  • Built the NSF Maestro platform for collecting behavioral data on spatial reasoning tasks at scale.

SDE Intern @ Amazon Web Services [Jun 2022 - Sep 2022]

  • Designed and implemented an automated EBS Snapshot cleanup pipeline in Python for cloud storage lifecycle management.
  • Optimized data partitioning, load balancing, and MapReduce strategies: 50% runtime reduction and 60% cost savings.

Software Engineer @ Samsung Research Institute [Jul 2017 - Sep 2020]

  • Led development of 4G and 5G cellular dongle interfaces for Samsung TV; integrated MBIM and QMI modem protocols, reducing network connection time by 83%.
  • Shipped dongle interface at CES 2020. Co-inventor on a US patent for cellular dongle interface technology.

Publications

Maestro: A Gamified Platform for Teaching AI Robustness [AAAI EAAI 2023]

Margarita Geleta, Jiacen Xu, Manikanta Loya, Junlin Wang, Sameer Singh, Zhou Li, Sergio Gago-Masague

  • Designed and developed open-source game-based platform for robust AI education, exposing students to adversarial ML challenges in a competitive programming environment.
  • Evaluated with 147 undergraduate students across two college courses; leaderboard identified as a key driver of engagement and learning.

Projects

Multi-Image Generation using Cycle GAN [Mar 2022 - June 2022]

  • Designed and implemented Augmented Cycle GAN to learn and generate many to many mappings of two domain using noise.
  • Injected noise acts as latent variable controlling image generation and varying it produces different versions of single image.
  • Analyzed performance of the system on Edges2Shoes, Night2Day, CelebA and FERDB datasets
  • Code is available in manikanta-72/multi-image-generation repository

Distributed Multi-Room Chat Application [Jan 2023 - Mar 2023]

  • Designed and implemented a multi-room chat web application in distributed environment and deployed it in AWS EC2 instances.
  • Publish and Subscribe architecture was implemented using Apache Kafka as the messaging middleware.
  • Java-spring boot and React were used as backend and frontend.