Sai Charitha Akula

Email / GitHub / LinkedIn / Google Scholar / Beyond AI / Life OS

I am a Senior Research Scientist at Meta Superintelligence Labs, where I work on multimodal AI systems for visual perception, reasoning, and reliable assistance. My research spans data curation, visual representations, post-training, and evaluation across text, images, and video. Prior to Meta, I was part of Millennium's four-person core AI team, reporting to the Head of AI. There, I built enterprise-grade RAG systems and agentic workflows from the ground up and conducted research on retrieval relevance. Before that, I researched vision-centric multimodal models and diffusion models at NYU's CILVR Lab under Prof. Saining Xie. I previously spent six years as a Quantitative Researcher at Goldman Sachs' Interest Rates desk, where I built models for automated trading, pricing, and risk management.

I completed my master's in Computer Science from NYU (Courant) and bachelor's in Computer Science (Honours) from Indian Institute of Technology, Bombay.

I have served as a reviewer for NeurIPS, CVPR, and ECCV.

Outside of research, you'll usually find me exploring new cafés, nature spots, and scents, dancing, staying active, or trying out something new!

Publications

ReconcileBench: Benchmarking State Reconciliation in Persistent Scientific Agents

NeurIPS 2026 Workshop on Agentic AI for Biological Discovery

Sai Charitha Akula, Chanukya Vardhan Gujjula.

A benchmark for whether persistent scientific agents can keep their workspaces consistent after evidence or analysis policies change. It tests reconciliation across data, code, figures, reports, and decisions in analytical, health, and RNA-seq projects, revealing how local repairs can leave downstream scientific records stale.

Oral Paper PDF Code Benchmark & data Coming soon

Selected for Oral Presentation at the NeurIPS 2026 Workshop on Agentic AI for Biological Discovery. View accepted oral papers.
LogT: Logically Think with Images for Visual Search

NeurIPS, 2026

Yanjun Fu, Quanzeng You, Jiadong Guo, Yujie Lu, Sai Charitha Akula, Haiping Wu, Miao Liu, Sanghamitra Dutta, Lu Yuan.

An automated data engine that teaches vision-language models to reason logically and use a zoom-in tool for visual search. LogT's supervised fine-tuning alone outperforms prior SFT + RL methods with 3.1× less data and over 120× fewer GPU hours; a simple RL stage extends the lead to 9.59 points across seven test sets.

Paper PDF Code Coming soon
Cambrian-1: A fully open, vision-centric exploration of multimodal LLMs

NeurIPS, 2024

Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, Saining Xie.

Oral Website Paper Code Datasets Benchmark

Selected for Oral Presentation (1.8%) at NeurIPS 2024
A Few Adversarial Tokens Can Break Vision Transformers

CVPR Workshop on Adversarial ML in CV, 2023

Sai Charitha Akula*, Ameya Joshi*, Gauri Jagatap, Chinmay Hegde.

Paper Code

Experience

Meta AI — Senior Research Scientist, Superintelligence Labs (2025–Present)
Millennium — AI Research Engineer, Core AI (Founding team of 4 · 0→1 AI initiatives · Reported to Head of AI) (2024–2025)
CILVR Lab, NYU — Deep Learning Researcher (2023–2024)
Goldman Sachs — Quantitative Research & Engineering (Analyst → Associate → Vice President, Interest Rates) (2017–2023)

Selected Projects

Efficient Group Preference Alignment in LLMs

Deep Reinforcement Learning, NLP, DPO, PPO

Developing a two-stage alignment framework for LLMs, exploring methodologies like "Model-agnostic Aligners + Adapters" and "PPO + DPO" to address common preferences initially, then fine-tune for group-specific preferences efficiently.

Paper Code Slides
Video Moment Localization

NLP, CV, Temporal Video Understanding

Added iterative attention to enhance query & video representation integration, alongside multi-level cross-modal interaction and content-boundary-moment interaction. Achieved a 3% R@1 IOU=0.5 improvement on Charades-STA over our baseline, SMIN.

Paper Code Slides
Perceiver-Resampler Cross-Attention Captioning

Natural Language Processing, Computer Vision

Developed 3-module architecture: frozen CLIP & GPT2 LLM, and trainable bridge with Perceiver IO (generates context for LLM fed via gated cross attention) & mapping network (aligns CLIP-LLM). Achieved 31.21 BLEU-4 score on COCO val set.

Paper Code Slides
Multi-Modal Federated Learning in Health Care

Medical Image Processing, TensorFlow, PyTorch Lightning

Dual-encoder model using ResNet-50 and ClinicalBERT, pretrained on MIMIC-CXR image–report pairs to learn aligned visual-text features, then finetuned on CheXpert for chest disease classification via Federated Learning. Evaluated the effect of varying client counts and IID/non-IID data distributions (volume, features, and labels).

Paper Code Slides
Market Flux Finder

Cloud and Machine Learning, LLMs, RAG, Docker, Kubernetes, GCP, gRPC

Built a financial RAG browser using LLMs to extract query-specific data and summarize insights. Used Cloud SQL with PostgreSQL (pgvector extension) for data storage and deployed microservices on GKE cluster for scalable serving; improved latency with GPU nodes and gRPC, and employed Prometheus for monitoring.

Paper Code Slides
Semi-Supervised Object Detection

Computer Vision, Semi Supervised Learning, Distributed Training, HPC

Pretrained ResNet-50 with Barlow Twins on 512k unlabeled data, used as FasterRCNN backbone, finetuned on 30k labeled data.

Paper