Nov 2025 – Nov 2025 · 0 mo

• I talk to ChatGPT about things - version 2.0 post GPT5
• Prompt Engineering Conf

Oct 2025 – Oct 2025 · 0 mo

• Limits of Imagination - an open source journey

Oct 2025 – Oct 2025 · 0 mo

• Bad Demos in the Building

Sep 2025 – Sep 2025 · 0 mo

• Talk on Athena as a platform in Python

May 2025 – May 2025 · 0 mo · Seattle, Washington, United States · Remote

• The Limits of Imagination : An Open Source journey
• My Pycon talk bio for a decade or so said these words : "I do fun experiments, sometimes social with data, python and language models because I believe as do most multinational companies now, that the human condition is encoded in language, just as science in math, and it in inevitable that one day, we will be using computers to help us linguistically as they do mathematically." in one way or the other - This has come true in the last 2/3 years. #Vision
• https://www.youtube.com/watch?v=mBOzULTcWlg

May 2024 – May 2024 · 0 mo · Pittsburgh, Pennsylvania, United States · On-site

• Lightning talk : Only Bad Demos in the Building
• What do forensic artists, 90s fashion, and 300 travel photos from Chennai have in common? They all became part of my generative AI experiments—and most of them didn't go quite as planned.
• At PyCon US 2024, I walked through a series of offbeat demos built with open-source transformer models—from trying to write my own travelogue with AI, to creating a "Clueless"-inspired fashion matcher, to figuring out how to generate synthetic literary datasets without duplicating the internet. The tooling came from Hugging Face, DreamStudio, Bing’s Image Creator, and a fair bit of Python duct tape.
• The talk wasn’t about perfection—it was about possibility. About asking questions like:
• Will forensic portrait artists keep their jobs?
• Can AI write stories you’d actually read?
• What happens when you trust LLMs with your wardrobe?
• Whether these demos were successes or “beautiful failures,” I’ll let the audience decide. Because when you experiment in public, the real win is what you learn when things don’t work.
• 🧠 Topics: Generative AI, NLP, synthetic data, open-source tools, creative coding
• 🧪 Demo-heavy. Opinionated. Human-first.
• #PyConUS #LLMs #GenerativeAI #Speaker #SyntheticData #Python

Mar 2024 – Mar 2024 · 0 mo · Seattle, Washington, United States · Remote

• The Story of SVMs
• upport Vector Machines (SVMs) have long been a pillar of classical ML—but Support Vector Clustering (SVC)? That’s where the math gets wild.
• In this talk, I explored the Hilbert-space-level magic behind SVC: mapping data to high-dimensional feature space via Gaussian kernels, defining minimal enclosing spheres, and projecting nonlinear boundaries back into human-readable clusters.
• We went deep into:
• Why SVM principles can be applied to unsupervised learning
• How kernel tricks let us cluster without assumptions about shape or number
• The importance of scale (q) and soft margin (C) in separating overlapping, noisy clusters
• When SVC outperforms traditional methods—like PCA-enhanced clustering on flower classification tasks
• It was mathy. It was visual. And it showed how the same embeddings we use in neural nets share deep roots with classic optimization theory.
• If you’ve ever wondered what Hilbert space really does or how to cluster in chaos, this was the tour.
• 🧠 Topics: SVMs, kernel methods, clustering, mathematical intuition
• 📚 Based on the foundational paper by Ben-Hur, Horn, Siegelmann, Vapnik
• 🖇️ Slides on request
• #SVM #MachineLearning #KernelTrick #Clustering #Speaker #Geekle

Mar 2024 – Mar 2024 · 0 mo · Seattle, Washington, United States · Remote

• Create your own data
• We expect LLMs to be factual. But what if we use them where accuracy doesn’t matter?
• In Create Your Own Data, I explored meaningful use cases for LLMs that don’t rely on reliability—but instead leverage what models already know: language structure, pattern recognition, semantics, and multi-modal reasoning.
• This talk walked through:
• 🌀 Building a tautology generator—synthetic data where self-consistency > truth
• 🧳 Using LLMs to caption real travel photos, not for facts, but for creativity and mood
• ⚙️ Techniques like distillation, fine-tuning, and how to reuse large models to build small ones
• We talked about when LLMs are better as structure generators, not search engines. And why that unlocks a quieter, more useful corner of the generative space—less hype, more intent.
• 🧠 Topics: Synthetic data, semantic use cases, distillation pipelines
• 🧰 Tools: Hugging Face, GPT-3.5, Keras, Stability AI, Bing Image Creator
• 📑 Slides: Link
• #LLMs #SyntheticData #Distillation #CreativeAI #Geekle #Speaker #TravelTech #NLP

Sep 2023 – Sep 2023 · 0 mo · Tallinn, Harjumaa, Estonia · On-site

• Are your friends bullshitting? An LLM story
• Ever had a friend say one thing… then say the opposite five minutes later? I have. So I built a model to catch them in the act.
• In this talk, I showed how NLP can detect contradictions in everyday language using two embedding models—BERT and GPT-2—on top of the SNLI dataset. We explored:
• How to model contradictions and entailments with sentence pairs
• Why this is key to LLM coherence, especially when hallucinations are common
• How contradiction models can help filter fake or misleading news
• The best part? You can use these same tools to build a lightweight bullshit detector of your own.
• 🧠 Topics: NLP, contradiction modeling, embeddings, LLM coherence, misinformation filtering
• 📦 Models: BERT, GPT-2, SNLI
• 👥 Audience Level: Intermediate
• 🧪 Track: PyData: ML & Stats
• #PyCon #LLMs #NLP #ContradictionDetection #FakeNews #BERT #GPT2 #Speaker

May 2023 – May 2023 · 0 mo · Vilnius, Vilniaus, Lithuania · On-site

• I talk to ChatGPT about things
• ChatGPT has captured the imagination of millions and sparked conversations about whether it is truly an Artificial General Intelligence (AGI). In this talk, we put ChatGPT to the test by asking it a series of riddles and challenging it with tasks that expose its limitations. I will provide a brief introduction to Large Language Models (LLMs), their training paradigms, and current shortcomings. Through slides and live examples, we will explore why ChatGPT struggles with contextual understanding, verbal math, and certain conceptual knowledge like Venn diagrams or sensory metaphors such as the sound of an egg cracking. Finally, we will discuss what these limitations reveal about the next generation of conversational AI models and compare ChatGPT’s performance to human intuition to better understand the boundaries of AGI today.

Jul 2022 – Jul 2022 · 0 mo

• Is the News Media Actually Polarized — or Are We Conditioned to Think It Is?
• This talk started as an NLP experiment that failed to prove what I thought it would — and ended up revealing something more important.
• I used neural networks and media datasets (including COVID coverage, headlines, and fake news satire) to test whether mainstream news media could actually polarize a model. The results surprised me: most of the bias we attribute to major publications might not stem from the articles themselves.
• It opened up a bigger question — are we importing social media polarization into our perception of the press?
• From headline sentiment to topic-induced bias (yes, “Trump” had an effect), this was a deep dive into framing, perception, and the limits of machine learning in analyzing social issues.
• 🧠 As language becomes the lens through which we measure society, this intersection of NLP and policy is only going to grow.
• Let’s talk about what is measurable — and what might still just be in our heads.
• #NLP #AI #MachineLearning #BiasDetection #MediaLiteracy #PolicyTech #DataScience #PyCon #TechForGood #LanguageModels

Jun 2022 – Jul 2025 · 3 yrs 1 mo · Redmond, Washington, United States · On-site

• Worked for the Viva Topics Platform team, integrated activities mined by a combination of curation and knowledge graph/template matching procedures - with a scaling factor of 6 million documents per tenant, improved design to reduce call volume by a factor of 3 from the Sharepoint UI to the preload layer. Implemented an auto-bot to improve MAU and encourage a product habit across multiple systems. Integrated GPT3/GPT3.5 with prompting techniques and information sources from various systems to integrate Topics as a service into Copilot offerings. From scaling Topics to 9B+ docs and standardizing OneDrive links, to prototyping Copilot features and securing SharePoint APIs, my journey at Microsoft has been multifaceted. Through bold experimentation, rigorous system thinking, and generous collaboration, I’ve helped evolve product surfaces, backend systems, and the culture behind them.

Jan 2022 – May 2022 · 4 mos · Worcester County, Massachusetts, United States · Remote

• Creates a database to simplify Covid19 jargon using Machine Translation.
• 🧪 How do we quantify and crowdsource language simplification for public health?
• During the EasyCovid19 project, I explored methods to simplify COVID-related terminology using a mix of NLP research, crowdsourcing infrastructure, and UX design inspired by Duolingo.
• Our aim: Make COVID-19 information accessible to everyone — especially non-native speakers, cognitively diverse users, and individuals without scientific backgrounds.
• What we worked on:
• 🔹 Designing and deploying a crowdsourcing app for term/sentence simplification (easycovid19.app)
• 🔹 Analyzing sentence complexity via readability metrics (Flesch-Kincaid, SMOG, Fog Index, etc.)
• 🔹 Reviewing NLP literature on stylistic paraphrasing, sentence simplification, and feature embeddings
• 🔹 Exploring methods to measure agreement across user-generated simplifications (e.g., Cohen’s Kappa)
• 🔹 Planning chatbot interactions using Zammo to guide contributors with in-line grammar/simplicity checks
• 🔹 Investigating how idioms, compound structures, and scientific terms contribute to sentence opacity
• We're now looking into using annotated corpora, structure-focused embeddings, and feedback loops to improve simplification consensus at scale.
• This project sits at the intersection of language accessibility, crowd-led NLP annotation, and computational social good — and there’s so much more to explore.
• 📎 Learn more about the project: https://easycovid19.org/project
• #NLP #AI4Good #SentenceSimplification #Crowdsourcing #PublicHealthTech #UXForAI #Accessibility #LanguageModels #CivicTech #ComputationalLinguistics

Jan 2022 – May 2022 · 4 mos · Amherst, Massachusetts, United States · On-site

• Worked with integrating hardware sensors with software to automate experiments.

Jan 2022 – May 2022 · 4 mos · Amherst, Massachusetts, United States · On-site

• 🧠 Built a React.js-based web application to support behavioral research on stereotype-driven information retention in children — in collaboration with the Behavioral Sciences Department at UMass Amherst.
• The goal: Move a traditionally in-person psychological study online, while preserving experimental fidelity.
• Key features included:
• 🔹 An interactive “spot-the-difference” image test designed to surface implicit stereotype recall
• 🔹 Logic for capturing visual attention bias and response patterns
• 🔹 Modular design for experiment reproducibility and deployment at scale
• This work enabled the lab to conduct large-scale, remote behavioral studies — bridging experimental psychology and modern web infrastructure.
• #ReactJS #BehavioralScience #HumanCenteredDesign #PsychTech #UXResearch #StereotypeBias #FrontendEngineering #CognitiveScience #RemoteResearch #UMassAmherst

Oct 2021 – Oct 2021 · 0 mo · Sweden

• Is the News Media Actually Polarized — or Are We Conditioned to Think It Is?
• This talk started as an NLP experiment that failed to prove what I thought it would — and ended up revealing something more important.
• I used neural networks and media datasets (including COVID coverage, headlines, and fake news satire) to test whether mainstream news media could actually polarize a model. The results surprised me: most of the bias we attribute to major publications might not stem from the articles themselves.
• It opened up a bigger question — are we importing social media polarization into our perception of the press?
• From headline sentiment to topic-induced bias (yes, “Trump” had an effect), this was a deep dive into framing, perception, and the limits of machine learning in analyzing social issues.
• 🧠 As language becomes the lens through which we measure society, this intersection of NLP and policy is only going to grow.
• Let’s talk about what is measurable — and what might still just be in our heads.
• #NLP #AI #MachineLearning #BiasDetection #MediaLiteracy #PolicyTech #DataScience #PyCon #TechForGood #LanguageModels

Oct 2021 – Oct 2021 · 0 mo · Hong Kong SAR

• Is the News Media Actually Polarized — or Are We Conditioned to Think It Is?
• This talk started as an NLP experiment that failed to prove what I thought it would — and ended up revealing something more important.
• I used neural networks and media datasets (including COVID coverage, headlines, and fake news satire) to test whether mainstream news media could actually polarize a model. The results surprised me: most of the bias we attribute to major publications might not stem from the articles themselves.
• It opened up a bigger question — are we importing social media polarization into our perception of the press?
• From headline sentiment to topic-induced bias (yes, “Trump” had an effect), this was a deep dive into framing, perception, and the limits of machine learning in analyzing social issues.
• 🧠 As language becomes the lens through which we measure society, this intersection of NLP and policy is only going to grow.
• Let’s talk about what is measurable — and what might still just be in our heads.
• #NLP #AI #MachineLearning #BiasDetection #MediaLiteracy #PolicyTech #DataScience #PyCon #TechForGood #LanguageModels

May 2021 – Dec 2021 · 7 mos · San Diego Metropolitan Area

• As part of the TurboTax Live team at Intuit, I contributed to building and modernizing key user-facing experiences that connect millions with live tax experts. A core focus was the migration from class-based to functional React components with Hooks, improving maintainability and performance across the codebase.
• 🔹 Refactored critical components using useState, useEffect, and custom hooks
• 🔹 Ensured accessibility and UI responsiveness across mobile, tablet, and desktop
• 🔹 Tested rigorously across various devices and edge-case user scenarios to ensure consistent performance and compliance
• 🔹 Collaborated with product, design, and QA to roll out updates during peak tax season
• 🔹 Improved code structure for faster iteration and reduced regression risk
• This experience sharpened my ability to build robust, scalable frontend systems in high-stakes, regulated environments, with a user base spanning devices and tax complexities.
• #ReactJS #Intuit #TurboTaxLive #FrontendEngineering #CrossDeviceTesting #FunctionalComponents #Hooks #Fintech #UXEngineering #Accessibility #PerformanceEngineering

Feb 2021 – May 2021 · 3 mos · Amherst, Massachusetts, United States · On-site

• 🖥️ UI Developer | NSF-backed Research Project – Colorado State University
• Built a web-based annotation interface used by ornithologists to label and analyze migratory bird behavior as part of an NSF-funded research effort. The interface streamlined large-scale data annotation tasks critical for ecological modeling and long-term species tracking.
• Key Contributions:
• 🔹 Designed and implemented a custom UI for efficient annotation and review
• 🔹 Integrated backend systems to support real-time data saving and user sessions
• 🔹 Optimized for usability by field researchers and non-technical domain experts
• 🔹 Collaborated closely with ecologists and research scientists at Colorado State
• 🔹 Aligned development milestones with NSF project timelines and compliance needs
• This work brought together scientific collaboration, ecological research, and practical interface design to support data-driven conservation science.
• #Ornithology #NSF #UIEngineering #AnnotationTools #ScientificSoftware #ReactJS #EcologyTech #UXDesign #ResearchSoftware #DataLabeling #MigratoryBirds

Feb 2021 – May 2021 · 3 mos · Amherst, Massachusetts, United States · Remote

• Developed a time-aligned dataset for Relation Extraction by grounding state-change events from Wikidata to NYT Corpus articles, helping denoise distant supervision signals by incorporating temporal awareness.
• 🔍 Key Contributions:
• Designed and implemented a full data processing pipeline aligning entities, relations, and timestamps between Wikidata and the NYT Corpus
• Modified and extended Facebook’s BLINK for batch-mode Named Entity Linking; contributed a codebase feature to improve annotation scalability
• Developed time-window filters to retain only temporally relevant entity mentions, boosting precision in relation extraction tasks
• Evaluated dataset quality using a RoBERTa-based RE model, tested against WebRED benchmark
• Handled 500GB+ knowledge graph data and 1.8M NYT articles; used scalable join strategies, entity disambiguation, and NoSQL optimization
• Authored a paper proposing our method; plans for open-sourcing BLINK extension
• 💡 This project bridges temporal reasoning, entity linking, and automated knowledge base curation—pushing state-of-the-art techniques toward cleaner and more context-aware extractions.
• #NLP #TemporalGrounding #EntityLinking #KnowledgeGraph #RelationExtraction #FacebookBLINK #Bloomberg #UMassAmherst #Wikidata #NYTCorpus #OpenSourceNLP #Python #RoBERTa

Jul 2020 – Jul 2020 · 0 mo · Ohio, United States

• Are We Conditioning Our Children to Accept Biases?
• Thunder Talk
• Language being the center of human interactions can be used affirmatively or destructively. It’s usage forms the basis of our societies, our stories and our narratives. When children grow up in a world where the stories consistently entitle them or belittle them, they will internalize these beliefs without question. NLP considerably automates the analysis that is paramount to setting the world right when it comes to these societal beliefs. NLP has the potential to be an excellent equalizer. Most children pick up notions from their educational texts and the rules at their educational institute. Thus, analyzing, identifying and eliminating cultural biases in the literature used to educate would be among the first steps to fostering an equal world. This presentation has an interdisciplinary basis, borrowing concepts from psychology, linguistics, literature, statistics and computer science.

Dec 2019 – Jun 2020 · 6 mos · Mumbai Metropolitan Region

• Force For Good (Antarang Foundation) : Volunteers with the internal tech-volunteer organization. Started off as the experienced engineer to take architectural directions and mentor junior members of the team, then took on the role of the Client Lead. Stack was Spring Boot, Angular 6, SQL and Azure cloud services.
• Managed a team of three engineers and acted as the liaison between the client NGO and the team. Successfully delivered four core features of the application including an exceptional feature that was added because of the Covid-19 pandemic.

Oct 2019 – Oct 2019 · 0 mo · Johannesburg Area, South Africa

• https://za.pycon.org/talks/24-detecting-propaganda-in-fake-news-using-natural-language-processing/
• In the age of social media, fake news isn’t just misinformation—it’s weaponized narratives. This project explores a technical approach to detecting fake news and identifying propaganda by combining keyword extraction, entity resolution, and event-based sentiment modeling.
• 🔎 How it works:
• Claim Verification:
• → Extracts high-confidence keywords using RAKE (Rapid Automatic Keyword Extraction).
• → Uses those keywords to query trusted news aggregator APIs.
• → Applies SpaCy’s similarity functions to evaluate alignment between a given article and retrieved mainstream content.
• Propaganda Detection:
• → Implements the Path Model of Blame, parsing sentences for patterns of blame/praise using POS tagging + regex rules.
• → Flags emotional manipulations like fear, outrage, sympathy, and detects mentions of political actors and narrative spin.
• 🧠 What’s unique?
• Detects not only misinformation but intentional persuasion strategies.
• Bridges NLP, information retrieval, and political linguistics.
• Quantifies textual manipulation using event-agent detection and threshold-based scoring.
• 📈 Results:
• While performance varies across topics, initial experiments showed up to 100% match for well-known cases—highlighting the promise of temporal + rhetorical alignment as a fake news detection strategy.
• 💡 Why it matters:
• Mob violence, election interference, and public manipulation are not just societal problems—they’re solvable technical problems.
• https://www.itweb.co.za/article/open-source-community-should-help-fight-fake-news/KPNG8v8d96Rv4mwD

Oct 2018 – Jan 2021 · 2 yrs 3 mos · Mumbai Area, India

• Works for Credit in the Corporate and Investment Bank LOB.
• Contributed to the optimization and parallel processing of the batch components that run to compute the End-of-day reconciliation to reduce the processing time by half.
• Deployed and delivered the first Feature Release for the Credit Derivatives space by creating deal-model classes, trade action APIs and trade matching infrastructure in python. Is currently working on the Broker clearing flow model and the Sales allocations workflow model for the next feature release. Implemented TDD, Agile and full scale integration testing. Is the SME for this component.
• Improved scale by introducing a tornado based load balancer REST service interface to the Venue services like Bloomberg etc.
• Works on an object based No-SQL in-house database called Athena that has server locations around the world and syncs data with respect to all the instances. The classes and APIs are written as wrappers and layers on top of the database abstraction layer.
• Supports multiple legacy applications, optimized the architectural direction for some of the reporting feed infrastructure to improve the queue management/ latency from peak volume of 300 requests/week to about 10 requests/week.

Oct 2018 – Oct 2018 · 0 mo · Hyderabad Area, India

• Proposal: https://in.pycon.org/cfp/2018/proposals/using-nlp-to-demystify-terms-and-conditions-and-summarize-the-contents~b2Aza/
• About a month ago, my inbox was flooded with emails starting with “We’ve updated our Terms and Conditions…” Like most people, I skipped reading them. And I work in financial services—legalese should be my cup of tea.
• But here’s the truth: even technically proficient users often bypass these documents because they’re dense, ambiguous, and intentionally obscure.
• This talk introduces a Python-based NLP engine that breaks down legal agreements into digestible summaries—highlighting the pros, cons, and privacy risks associated with any Terms & Conditions document.
• 💡 Why this matters:
• In an age of rampant data breaches and opaque privacy practices, it’s no longer acceptable to click “I Agree” blindly. From political manipulation via social media to financial fraud, the implications of uninformed consent are massive.
• 🔍 What the system does:
• Extracts Gives and Takes from any legal agreement—identifying what the user is giving (e.g., data, fees, obligations) vs. what they are getting (services, access, guarantees).
• Quantifies tradeoffs using NLP metrics and semantic weighting.
• Flags suspicious or sensitive clauses, especially those related to data sharing, arbitration, or third-party liability.
• 🧠 Technical components include:
• Rule-based phrase extraction and legal pattern matching
• Semantic similarity scoring to detect subtle coercive clauses
• A scoring system for summarizing contracts by risk level
• 🙋 Who this is for:
• Anyone interested in digital rights, privacy, or automated legal understanding. No prior legal expertise is required—just a curiosity to understand how technology can make contracts transparent and ethical.

May 2016 – Jul 2016 · 2 mos · India

• Developing software solutions
• Worked with: Spring MVC, Backbone.js, Angular.js
• Modifying and correcting according to the requirements for a live-in-use webapp
• Developed an entire back-end to Add/Edit/Delete by envisioning suitable modals and connecting an Angular response on the front-end.
• Modified the front end so that it would be responsive on a mobile phone using Backbone
• Tidied up the front end, improved the overall CSS and changed the elements to standard Bootstrap elements.

– Present

May 2016 – Nov 2016 · 6 mos · India

• Handled three portfolios namely, Guest Lectures, Exhibits and Attractions. Hand picked and personally mentored a team of 15 to work as sub coordinators and workforce for the same departments.
• Guest Lectures: Curated a TEDx level event as a division of the Technozion 2016. The line up which was well attended and publicized consisted of scientists, entrepreneurs and experts spanning artificial intelligence, communication networks, marketing and start-up founders.
• Exhibits and Attractions: From a zilch in 2015 to 10+ attractions and exhibit events were conducted as a part of Technozion 2016. Exhibits consisted of social initiative, school children led exhibitions, Virtual Reality and 3D printing technology exhibits. The attractions consisted of a Martian Augmented Reality experience, A gaming competition, Holographic experience and Van De Graaff generators.
• Collaborated with INK, which in turn organised the first ever Makeathon at NITW with the Lakshya Foundation.

Dec 2015 – Feb 2016 · 2 mos · NITW

• Made a self playing version of a mobile piano app using image processing in MATLAB.
• https://github.com/GameAutomators/eBook-Source

Jul 2015 – May 2016 · 10 mos

• Development of Applications and fulfillment of user stories on Freecodecamp.com .
• Frameworks and Resources:
• Angular.js, Node.js, MongoDB, JQuery.js

Jan 2015 – Jan 2016 · 1 yr · NITW Student Branch

• Women in Engineering Secretary.
• As a member of IEEE, I motivated girls to become a part of it's global network and achieved a record-breaking membership count from their side. A large part of the WIE summit for school going girls in Warangal was taken care of in my tenure.
• Mentored projects, encouraged participation and collaboration among the members in various events in the home student branch as well as others in the Hyderabad zone. Was a part of the conduction team for the Entrepreneurship summit and MATLAB workshop organized by the NITW Student Branch.

Aug 2013 – Apr 2014 · 8 mos

• A contributor to a blogging platform as an outreach branch of the Honeybee Chapter, NITW. Served as a medium of encouragement to all the budding entrepreneurs in the college.
• The articles covered deconstruction of interesting TED talks, social and grass root innovators and their innovations and interactions with dignitaries in the same field.