Smriti Singh
I build agentic AI systems for production, with a focus on making them reliable, interpretable, and fair.
What I Build
Production systems where research-informed engineering makes the difference.
End-to-End Agentic Pipelines
Designed and led an agentic AI system for generating institutional-grade research reports on large-cap equities, projected to drive significant revenue impact at scale. Full ownership from architecture to deployment.
Hallucination-Resistant Fact-Checking
Built a student-teacher LLM chain for automated fact verification in AI-generated financial content, reducing hallucination risk and enabling trust in downstream applications.
NL-to-SQL Evaluation Framework
Developed a rigorous evaluation framework for measuring LLM performance on complex natural language to SQL tasks, including semantic correctness, execution accuracy, and failure mode analysis.
Fine-Tuning with Automatic Feedback Loops
Built an automatic feedback pipeline for iterative LLM fine-tuning on domain-specific financial tasks. Achieved a 33% boost in model performance through RLHF-informed iteration.
Custom LLM for Structured Content Generation
Led end-to-end ideation and implementation of a custom LLM that transforms dense financial articles into concise, insight-driven content. Shipped 200+ officially published pieces within 3 weeks of launch.
Research That Informs Engineering
My work on bias detection, emotional intelligence in LLMs, and moral reasoning directly shapes how I approach evaluation, failure analysis, and trust in the systems I build.
About Me
I'm a Machine Learning Research Engineer with hands-on experience designing, building, and shipping agentic AI pipelines in production. At Zacks Investment Research, I lead ML system design and applied research for agentic workflows in finance, owning the full lifecycle from architecture and prototyping to stakeholder-driven iteration and deployment.
My engineering is shaped by my research. Work on bias, emotional intelligence in LLMs, and moral reasoning gives me a different lens when I'm building evaluation frameworks, debugging agent failures, or thinking about what it means for a model to be "reliable." I hold an M.S. in Computer Science from UT Austin (minor in Computational Linguistics, NSF-certified portfolio in Ethical AI), where I was advised by Dr. Jessy Li and worked with Dr. Raymond Mooney.
I believe that making AI systems genuinely useful in high-stakes domains requires confronting the hard questions around alignment, interpretability, and fairness. Not as separate concerns, but as engineering requirements. That conviction drives both what I build and what I study.
My Story
Ever since I was a child, I deeply believed in the power of language, technology and science to transform lives. Growing up, I grappled with whether I wanted to be a writer, an engineer, or a scientist. It wasn't until I discovered the field of natural language processing during my undergraduate, that I realized I could combine all three passions into a single career.
My journey into the world of AI began with baby steps. As time passed, I became increasingly interested in the ethical implications of AI systems and the importance of building models that are not only powerful but also aligned with human values. My research started out focused on fairness and safety, spanning topics like sexism detection, threat detection using NLP, and multimodal misogyny detection in memes. Over time, I began to realize that the real problem was not just the visible biases in models, but the underlying misalignment between AI objectives and human values. I also realised that any technology is only as good as the challenges it can solve effectively in high stake domains, and so I expanded my professional focus to applied ML engineering for finance.
These experiences shaped a clear mission for my work: to build AI systems that are not only intelligent, but emotionally aware, interpretable, and aligned with human intent. I believe that superintelligence without emotional intelligence risks amplifying harm rather than insight, and that interpretability is essential for ensuring safety, accountability, and trust. Equally important, AI must move beyond theoretical benchmarks to solve real problems at scale, particularly in high-stakes domains like finance. Rooted in the firm belief that technology without intentional inclusion is just sophisticated discrimination, my approach centers on developing safe, fair systems that solve real-world problems.
My perspective is shaped by both privilege and responsibility. I am the first woman in my family to pursue a career, made possible by parents who believe deeply in equality and consistently support my ambitions. I am especially conscious that I stand on opportunities my mother sacrificed, and that awareness informs a core principle of my work: technology must not compound harm for underprivileged or historically excluded communities. I am a strong advocate for women in STEM and for gender equity in education and professional spaces, not as an abstract ideal but as a necessary condition for building better systems. My father's career in technology was a defining influence, instilling both technical curiosity and a respect for disciplined engineering. In a field moving at unprecedented speed, I hope my work contributes to a more deliberate trajectory: one where progress toward superintelligence is matched by interpretability, emotional awareness, and a commitment to fairness and safety, and where others are inspired to build not just faster systems, but more responsible ones.
Technical Skills
Education
M.S. Computer Science
Minor in Computational Linguistics, NSF-certified portfolio in Ethical AI. Advised by Dr. Jessy Li; thesis published at NAACL 2024. Co-authored another paper with Dr. Raymond Mooney (ACL 2023).
B.Tech Information Technology
Minor in Big Data. First in cohort to land an ML Research offer in the Bay Area. Published research across sexism detection, child predator detection, and threat detection at top venues like ACL and IEEE.
Research
My research informs how I build. Bias detection work shapes my approach to LLM evaluation; interpretability work informs how I debug agentic failures.
Research Interests
Alignment × Interpretability
How interpretability techniques can be used to understand, evaluate, and improve alignment, making internal representations and failure modes transparent for safe deployment.
Emotional Intelligence in AI
Studying affect modeling and social context as alignment components, and how emotional awareness contributes to safer, more human-aligned decision-making in advanced systems.
Fairness & Bias Mitigation
Developing scalable methods to detect and mitigate bias in language and multimodal models, with approaches that promote fairness without compromising real-world performance.
Intelligence Evaluation
Examining how different architectures and training regimes converge to similar phenomena, and what this implies for measuring and stress-testing intelligence beyond narrow benchmarks.
Select Publications
Reasoning or Rhetoric? An Empirical Analysis of Moral Reasoning Explanations in Large Language Models
Smriti Singh, Aryan Kasat, Vinija Jain, Aman Chadha
Investigates whether LLMs engage in genuine moral reasoning or produce post-hoc rhetorical justification, with implications for alignment evaluation and trustworthy AI systems.
View Paper ↗Born With a Silver Spoon? Investigating Socioeconomic Bias in LLMs
Smriti Singh, Shuvam Keshari, Vinija Jain, Aman Chadha
Introduces SILVERSPOON, a 12,000-sample dataset for multifaceted analysis of socioeconomic bias. Demonstrates that state-of-the-art LLMs exhibit both explicit and implicit socioeconomic bias, compounded by intersecting gender and racial stereotypes.
View Paper ↗From Prejudice to Parity: A New Approach to Debiasing Large Language Model Word Embeddings
Smriti Singh, Aishik Rakshit, Shuvam Keshari, Vinija Jain, Aman Chadha
Proposes DeepSoftDebias, a neural soft-debiasing algorithm that outperforms state-of-the-art methods across gender, race, and religion bias benchmarks.
View Paper ↗Language Models (Mostly) Do Not Consider Emotion Triggers When Predicting Emotion
Smriti Singh, Cornelia Caragea, Junyi Jessy Li
Thesis work. Reveals that human-annotated emotion triggers are largely not considered salient by emotion prediction models, with implications for emotional intelligence and interpretability in LLMs.
View Paper ↗"Female Astronaut, because sandwiches won't make themselves up there!": Towards Multimodal Misogyny Detection
Smriti Singh, Amritha Haridasan, Raymond Mooney
Addresses the challenge of detecting misogyny in multimodal meme content, exploring the role of domain-specific pretraining in digital discourse analysis.
View Paper ↗Blog
Writing on AI safety, emergent behavior, and the big-picture questions in the race toward superintelligence. On Substack.
Early Signs of the Singularity? What Moltbook Reveals about the Future of AI
Over 150,000 AI agents signed up for their own social network in 72 hours, creating communities, debating philosophy, and discussing ways to communicate without human oversight. What does Moltbook reveal about multi-agent AI systems, alignment, and the governance challenges ahead?
Read More ↗Claude's New Constitution: Building the Guardrails for Safe Superintelligence
Understanding Anthropic's new approach to safety in superintelligent systems, with a focus on key takeaways and necessary next steps on enforcement.
Read More ↗Language as the Substrate of Intelligence: A Hypothesis on Emergent Cognition
A hypothesis on how language itself may be what drives the emergence of intelligent capabilities across species, backed by neuroscience and psychology.
Read More ↗Evolution, Equifinality and Emotional Intelligence: The Unsolved Mysteries in the Race for Super-Intelligence
A deep dive into some of the big-picture questions that need to be answered to ensure a safe future for an AI-driven society.
Read More ↗When Prediction Looks Like Purpose: Understanding the Paradoxical Emergent Behavior in LLMs
An exploration of how large language models exhibit goal-directed behavior despite being trained solely for next-word prediction, and what this means for AI alignment.
Read More ↗Media Features
NeurIPS 2025: Investigating Socioeconomic Bias in LLMs
Featured on the WiAIR podcast after presenting findings on socioeconomic bias in large language models and discussing their implications for responsible AI development.
Watch Video ↗The Urgency of Fairness and Interpretability in AI
A featured article explaining why fairness and interpretability are critical components of responsible AI development, and why these concerns cannot wait for later in the development cycle.
Read Article ↗Would an AI Judge Be Able to Efficiently Dispense Justice?
Research on the potential and challenges of using AI as judges in legal systems, featured in an article exploring AI decision-making in high-stakes institutional contexts.
Read Article ↗The Role of Domain-Specific Pretraining in Digital Discourse Analysis
Findings on the technical and social challenges of multimodal misogyny detection in memes, featured by UT Austin's CS department.
Read Article ↗Invited Talk: Using NLP to Prevent the Spread of Gendered Health Misinformation
Invited talk on gendered health misinformation and NLP techniques for detection and prevention at scale.
Watch Video ↗Invited Talk: Understanding How AI Can Be Used to Tackle Hate Speech Online
Analyzing the challenges and opportunities of using AI to detect and mitigate hate speech at scale across online platforms.
Watch Video ↗Invited Talk: Mental Health Matters! NLP Techniques in the Mental Health Domain
Investigating how NLP techniques can be leveraged to build applications for mental health support and early intervention.
Watch Video ↗Panel Discussion: The Future of AI in Healthcare
Invited panelist discussing the potential and challenges of using AI to improve healthcare outcomes, with a focus on fairness and responsible deployment.
Watch Video ↗Updates & More
Journey Highlights
2026
- "Reasoning or Rhetoric?" accepted to the Logical Reasoning in LLMs Workshop at ICLR 2026
- Invited as a reviewer to multiple ICLR 2026 workshops
- Founded the Women in AI Research Mentorship Lab, launching May 2026, with an advisory board including Aman Chadha (Apple Gen AI), Vinija Jain (Meta AI), and Sreyoshi Bhaduri
2025
- Paper accepted at NeurIPS 2025 (LLM Eval Workshop) on socioeconomic bias in LLMs
- Co-authored paper published at COLING 2025 main with a mentee
- Research presentation featured on Women in AI Research podcast
- Blog piece accepted at The AI Journal
2024
- Graduated from UT Austin with an M.S. in Computer Science
- Thesis accepted at NAACL 2024 main; attended first in-person conference
- Joined Zacks Investment Research as an ML Research Engineer
- Research featured in the New Scientist
2023
- Co-authored paper accepted to ACL (WOAH Workshop)
- Research featured in the UT Austin CS Newsletter
- First journal paper published at IEEE Access with a mentee
- ML Research Engineer internship at Esperanto Technologies, Mountain View
2022
- Joined UT Austin with full funding, after offers from UW, Georgia Tech, and UBC
- Only student in undergraduate cohort to land an ML Research offer in the Bay Area
- Invited talk at PyData Hamburg on Gendered Health Misinformation
- Received PyData Impact Scholar Award
2021
- First paper published at ACL SRW
- Won MITACS Global Research Award; research internship at Ryerson University, Toronto
- Received Grace Hopper Scholar Award and EmTech MIT Scholar Award
- Invited panelist at FruitPunch AI on AI in healthcare
2020
- Led team to win district-level Smart India Hackathon; qualified for finals
- Won Hub level of Google Hashcode
- Mentees and I published two papers at WinLP@EMNLP
2019
- Invited to talk on DNA Computing for AI at BITS Pilani
- Elected Vice Chair of ACM-W Manipal
2018
- Joined Manipal Institute of Technology for B.Tech in Information Technology
- Won best idea award at Manipal Hackathon for research on DNA computing for AI