Abbaas Alif Mohamed Nishar · Applied Scientist at Capital One
I build conversational AI and study where it goes wrong.
I fine-tune models for multi-turn conversations, build evaluation tools that explain why assistants fail, and develop tools to deploy GPU-backed models. My research covers video-language evaluation, neural decoding, and event cameras. I hold a Ph.D. in Computer Science from Georgia State University.
I work across the lifecycle of conversational AI: training models to route requests, evaluating what happens after routing, and building tools to operate models in production.
Multi-turn intent classification
I fine-tune language models to route requests that depend on earlier turns. Retrieval narrows the candidate intents, and constrained decoding limits predictions to those candidates. I evaluate failures on ambiguous requests, missing context, and overlapping labels.
Conversation evaluation & failure analysis
I built review tools that group recurring failures and separate a routing mistake from a task that remained unfinished even though the request was understood. That gives reviewers evidence for what to fix next.
Model deployment & operations
I built tools to deploy and monitor GPU-backed language models and route proposed operational changes for human approval.
Jul 2023–Jun 2025
Revelio Communications
Founding Engineer
Built the screen-to-camera pipeline for embedding data in video and recovering it on a smartphone. Worked across adaptive encoding and the two-stage neural decoder, including cases where the screen and camera were not synchronized.
Built and evaluated optical communication and sensing prototypes with event cameras. Developed a non-line-of-sight receiver using asynchronous pixel events and investigated how reflective surfaces and modulation affect signal recovery.
Developed image segmentation for home-inspection imagery and automated multi-GPU training.
Aug 2020–Aug 2021
Tata Consultancy Services
DevOps Engineer
Built retail forecasting dashboards and Jenkins CI/CD pipelines.
Dec 2019–Jun 2020
Tata Consultancy Services
Research Intern
Prototyped AR/VR and computer vision applications for clothing-pattern detection and text-to-avatar interfaces.
Selected work
Industry work
When the assistant understands but doesn’t help
I work on two parts of a conversation: recognizing what someone wants and checking whether the response helped them do it. I built evaluation tools that let reviewers distinguish a misunderstood request from an incomplete answer. That distinction points to two different fixes: change how the request is interpreted, or change what happens next.
Fictional example · Library assistant
Reader: “Can I renew my book?”
Assistant: “To renew a book, go to your account.”
Request understood
Yes
Renewal confirmed
No
The reply gives instructions; it does not establish that the book was renewed.
Video understanding
REVEAL: Did the model watch the video?
What happens when a familiar action plays backward? In REVEAL, my co-authors and I tested whether video-language models followed the action on screen or described what they expected to happen. Some models described the forward action even when the clip was reversed.
Published scope: Five controlled tests across 12 video-language models, covering failures such as weak cross-frame integration and reliance on language cues.
My co-authors and I built a way to embed information in video without adding a visible code to the picture. A standard smartphone camera captures the screen, and a neural decoder recovers the message even when the screen and camera are not synchronized.
Published result: Zero decoding errors across four videos at distances up to 2 m in direct-view tests.
A Framework for Classifying Applications from Raw Network Traffic Traces
IEEE SoutheastCon 2024 · Best Student Paper Award
Joint Optical Wireless Communication and Sensing Using Neuromorphic Cameras
ACM/IEEE IPSN 2024 · Third Place Best Poster
Patent
Machine learning based system and method for controlling residual artifacts in media contents to optimize user experience in real-time screen-to-camera communication environment
U.S. Patent 12,676,939 · 2026 (application US20250030811A1)