Meghna Allamudi
Building intelligent dialogue systems and multi-agent AI at the intersection of research and industry.
01. About Me
I'm a Research Engineer passionate about building AI systems that understand and engage in natural human dialogue. My work spans both cutting-edge research and large-scale production systems.
Currently, I'm a Software Engineer at Snapchat, building attribution systems that process ~80B daily interactions, while simultaneously pursuing my M.S. in Computer Science (ML Track) at Columbia University.
At Columbia, I research in Columbia NLP Lab | Dr. Zhou Yu, where I've architected multimodal negotiation AI systems with voice cloning and real-time analytics, deployed to Columbia Business School and other Ivy League communities.
I believe in the transformative power of AI agents. My side projects include a multi-agent murder mystery platform that reached 1,500+ users in one week, demonstrating how generalized agent networks can create engaging, hallucination-free content at scale.
Education
M.S. Computer Science
Columbia University • 2023-2026
ML Track (Part-Time)
B.S. Computer Science
Rose-Hulman • 2018-2022
02. Experience
Software Engineer, Monetization
@ SnapchatOct 2024 - Present • Seattle, WA- ▹Built attribution systems processing ~80B daily user interactions to determine conversions, informing $50B+ in annual advertiser spend decisions.
- ▹Architected cross-platform attribution reconciliation with Google Analytics & TripleWhale, unlocking $324M in incremental annual ad spend.
- ▹Designed end-to-end targeting optimization platform affecting 200+ of top 400 web advertisers, processing ~2M records in <10 minutes using distributed programming.
- ▹Shipped distributed caching service improving Ad Ranking model accuracy by filtering 97% of invalid requests, impacting ads seen by 400M+ daily active users.
- ▹Built AI-powered RCA agent automating advertiser issue diagnosis across petabytes of data; leading team of 6 engineers (L4-L5).
- ▹Designed production-grade attribution algorithm for GDPR cookie consent scenarios, recovering lost conversion data for top advertisers including VANS, Temu, Booking.com.
Software Engineer, Microsoft AI (MSN)
@ MicrosoftSept 2022 - Sept 2024 • Redmond, WA- ▹Owned end-to-end revenue pipeline for Microsoft News (MSN), the financial backbone determining publisher payouts and A/B test decisions across petabytes of interaction data.
- ▹Drove ~11% latency reduction in MSN's core data pipeline powering recommendation systems for 400M+ monthly active users.
- ▹Owned third-party content partner data infrastructure representing 8% of DAU (~30M+ users), building unified platform integrating external partner data with first-party telemetry.
- ▹Designed warm/cold-path data parity system in Spark; built self-service data API reducing time-to-insight from hours to minutes.
- ▹Architected production revenue attribution system ingesting data from ~20 advertising partners daily, reconciling cross-platform conversion events.
Software Engineer Intern
@ MicrosoftJune 2021 - Aug 2021 • Redmond, WA- ▹Worked on the Local News Team part of the News and Feeds group.
- ▹Created a production dashboard tracking user-engagement metrics for local news, including location-based and document-based (CTR by position) metrics.
- ▹Pre-processed multiple large datasets (9GB+ each), built data pipeline with scheduled automation, and integrated with PowerBI via Azure Data Lake Storage.
- ▹Presented finished project to Corporate Vice President of Web Experiences Team, Taroon Mandhana.
Software Engineer & PM Intern
@ MicrosoftMay 2020 - Aug 2020 • Redmond, WA- ▹Worked as front-end developer and PM for the Clarity Team (Bing Shopping/Commerce) - a web analytics tool for tracking customer usage.
- ▹Created production sharing workflow from scratch for users to share data with teammates and publicly; wrote design specs for major workflow aspects.
- ▹Built share modal platform, usage logging for success measurement, and public shareable link authentication bypass.
- ▹Presented project to GM for Bing Shopping/Commerce, Manish Mittal, and his directs.
Full-Stack Developer
@ Direct SupplyJune 2019 - March 2020 • Milwaukee, WI- ▹Worked on back-end and front-end of E-Commerce platform and Order Guide Management platform.
- ▹Built fully-functioning cross-platform app with 5 interns, presented to CEO and executive board (currently used by company).
- ▹Implemented production features including stored procedures, Swagger UI authentication, and order guide workflows.
03. Research
Research Intern
Columbia NLP Lab | Dr. Zhou Yu
March 2025 - Present
Competitively selected following successful submission of an AI agent debate demonstration using Arklex. Working directly with PhD researcher Ryan Shea on advanced negotiation AI systems.
- ▹Architected multi-issue negotiation AI encompassing automated mediation between competing agents and sophisticated feedback mechanisms for skill development.
- ▹Technical lead for Columbia Business School integration: designed and deployed comprehensive web platform enabling business students to engage with negotiation bots and receive real-time performance analytics.
- ▹Spearheading development of speech-enabled negotiation capabilities, implementing NLP pipelines that analyze verbal negotiation patterns and provide targeted feedback on communication effectiveness.
- ▹Achieved 95% accuracy for voice classification across Switchboard-1 and POM audio datasets.
- ▹Co-authoring paper with PhD researcher Ryan Shea; research directly supports Columbia Business School curriculum while advancing conversational AI for professional skill development.
Publications
Persuasive Dialogue Corpus: Graph-Based Approach Combining Persuader and Persuadee Perspectives
M. Allamudi, O. Scrivner
Proceedings of the Future Technologies Conference, 607-621, 2022
04. Projects
Cold Case Files
AI Multi-Agent Murder Mystery Platform
Architected a generalized multi-agent network with specialized sub-agents per case facet (evidence, suspects, witnesses, timeline) enabling cross-agent reasoning with zero hallucinations. Full-stack solo build.
AI Summarizer
Intelligent Document Summarization
Built an AI-powered summarization tool that processes documents and text to generate concise, accurate summaries using advanced language models.
Adaptive Persuasive Dialogue Agent
Meta-Learning for Dialogue Strategy
Applied MAML (Model-Agnostic Meta-Learning) to dialogue strategy selection, achieving 17-18% improvement over population-best baselines and 33% over random with only 3-5 adaptation examples.
Debate AI Agent
Arklex Framework Implementation
Created a sophisticated debate agent with dynamic rhetorical adaptation using RAG for knowledge retrieval and an Argument Classifier for classical rhetorical modes (logos, ethos, pathos).
Neo4J Persuasion Graph
Dynamic Strategy Adaptation
Built an AI agent using Neo4J knowledge graphs to model conversational state transitions, with sentiment analysis triggering strategy backtracking for self-optimizing dialogue.
Neuroscience-Inspired Noise Augmentation
Language Model Research
Research project applying neuroscience principles to LM architecture by introducing strategic noise during training, targeting attention mechanisms in BERT and GPT models.