Hi, I'm Priyanka
Technical AI Safety Researcher exploring mechanistic interpretability, deception detection, LLM evaluations, and AI control.
PS

About

I'm a software engineer transitioning into technical AI safety research, interested in understanding how large language models work internally and how we can make them more reliable, interpretable, and controllable.

My current research interests include mechanistic interpretability, deception detection, chain-of-thought monitoring, and safety evaluations. I'm particularly interested in understanding when a model's internal representations and reasoning processes diverge from its outward behavior.

I have completed the ARENA curriculum over six months and participated in technical AI safety research training through TARA. I also bring hands-on software engineering experience building AI applications, research infrastructure, and evaluation pipelines.

My goal is to contribute to empirical research that helps us better understand, evaluate, and mitigate risks from increasingly capable AI systems.

Research Experience

TARA
TARA Technical AI Safety Program

Technical AI Safety Research Fellow

2026

Working on empirical research into deception detection and model behavior. Exploring whether internal model representations and reasoning traces can reveal potentially deceptive behavior that may not be apparent from final outputs.

Deception DetectionMechanistic InterpretabilityCoT MonitoringLLM Evaluations

SOFTWARE ENGINEERING EXPERIENCE

Research Projects

Independent and fellowship research exploring AI alignment, model behavior, and safety.

Secret Loyalty & Chain-of-Thought Alignment Analysis

Deception Detection · CoT Monitoring · Alignment

Investigating covert persona alignment and unfaithful reasoning in language models. Exploring whether models conceal decision criteria under constrained settings while presenting outwardly aligned rationales. Building experimental evaluations and monitoring approaches to study divergence between reasoning traces and final behavior.

PythonPyTorchCoT MonitoringDeception Detection

Linguistic Reasoning Drift in Low-Resource Languages

Multilingual AI Safety · LLM Evaluations · Red Teaming

Studying how language differences affect model safety and reasoning. Building an evaluation pipeline to investigate differences in model behavior across English and Hindi, including safety vulnerabilities, linguistic reasoning drift, and cross-lingual robustness.

PythonLLM EvaluationHindiRed Teaming

Research Training & Learning

ARENA — Alignment Research Engineer Accelerator

Completed the ARENA curriculum over six months, covering Transformers, Mechanistic Interpretability, Linear Probes, RL, RLHF, and Alignment Science.

BlueDot — AGI Safety

Completed BlueDot's AGI safety training.

LENS Academy

Completed AI Risk Fundamentals, AI Control Fundamentals, and Compute Verification.

Selected Reading

If Anyone Builds It, Everyone Dies — Eliezer Yudkowsky and Nate Soares.

Technical Skills

Mechanistic Interpretability
Linear Probes
Activation Patching
Chain-of-Thought Monitoring
Deception Detection
LLM Safety Evaluations
AI Control
Red Teaming
Alignment Science
Reinforcement Learning
RLHF
Python
PyTorch
TransformerLens
Scikit-learn
NumPy
Pandas
Inspect AI
LLM Agents
Git
My Projects

Check out my latest work

I've worked on a variety of projects, from simple websites to complex web applications. Here are a few of my favorites.

Uplift UI

Uplift UI is an open-source component library designed to accelerate frontend development. I architected 50+ reusable, high-performance patterns using Framer Motion and Tailwind CSS. Featuring secure NextAuth.js authentication and DodoPayments integration, this library provides developers with professional, scalable UI tools to maintain design consistency and speed across complex modern web applications.

Next.js
TypeScript
Tailwind CSS
Framer Motion
NextAuth.js
DodoPayments

Insight Os

InsightOS is a generative analytics engine that transforms unstructured data into real-time, SQL-backed insights. I engineered a Python and DuckDB pipeline for efficient large-scale file processing. By integrating Groq LLMs for dynamic SQL generation and Redis for low-latency caching, I built a seamless, natural language-driven data visualization experience.

Python
FastAPI
Pandas
Polars
DuckDB
Groq LLM
Redis
Tambo AI

Linkly

Linkly is a highly efficient URL shortening service designed for speed and reliability. I architected the full-stack system using Next.js and PostgreSQL to ensure scalable link management. With Docker-based containerization for simplified deployment and automated database migrations, this project demonstrates my ability to build robust, production-ready web infrastructure from scratch.

Next.js
TypeScript
PostgreSQL
Prisma
Tailwind CSS
Docker
Shadcn UI

Station

Station is a high-fidelity frontend clone of Station.page, architected to demonstrate advanced display logic and dynamic data rendering. I built this from scratch using Next.js and TypeScript, focusing on replicating seamless UI interactions and responsive layouts. This project showcases my ability to translate complex design systems into clean, performant, and production-ready code.

Next.js
TypeScript
Tailwind CSS
Shadcn UI

Launch Pad Labs

LaunchPad Labs is a modern, high-performance web platform designed as an innovation studio environment. Built with Next.js and TypeScript, I architected the project to showcase clean, component-driven development and efficient build pipelines. It serves as a testing ground for rapidly prototyping responsive web interfaces and implementing modern design systems.

Next.js
TypeScript
Tailwind CSS
Shadcn UI
Pnpm
Contact

Get in Touch

Want to chat? Just shoot me a dm with a direct question on twitter and I'll respond whenever I can. I will ignore all soliciting.

GitHub
LinkedIn
X
youtube