I work on models of how people watch: systems that choose where to look, perceive the consequences, and are judged against real human behaviour rather than benchmark accuracy alone. I also build products.

Currently an MRes researcher at the UAL Creative Computing Institute with Tim J. Smith. Before that: MTech in Computer Science & Engineering (AI), AI R&D at Digital University Kerala, and a co-founder of an AI interview platform. I also make films.

London · rahulrajpv.vercel.app · LinkedIn · rahulrajpv.ai@gmail.com

pitute: the network for researchers
A place to find every scholarship, PhD, postdoc and research opportunity in the world, with identity drawn from academic infrastructure (ORCID, institutional email, the publication record). Live since October 2026.

• 90+ importers read funders, ministries and universities' own recruitment systems (Workday, Oracle, PageUp, Stonefish and others) across about 100 countries.
• Each member is matched by a hybrid scorer (dense embeddings, catalogue topics, field words). Eligibility is read as quoted facts from the official page and checked by deterministic rules; a fact whose quote is not on the page is dropped.
• One Next.js app with 88 Prisma models, 14 scheduled jobs and 3,505 unit tests.

JOY: a closed-loop model of how people watch film (MRes, UAL)
A closed-loop observer for film. It sees only a foveated percept, arbitrates between audiovisual salience and a vision-language model's reading of the scene, selects its own fixations with human-like timing, and perceives the world from wherever that choice left it. The main finding is a dissociation: a state-dependent policy that improves prediction of human gaze in 5 of 5 folds wins on only 3 of 8 clips once it controls the loop. Prediction and observation are different problems.

The system includes:

• a gaze-contingent scanpath loop with inhibition of return and CRISP fixation durations
• a LoRA-tuned Qwen2.5-VL that reads fixations out as coordinate tokens
• persistent character identity and speaker grounding (BCubed F 0.972 / 0.985)
• an engagement pipeline linking salience, event segmentation and goal inference
• models of affective and pupil response

Project page · site source · research code available on request

JOY Companion
A real-time, fully on-device companion that perceives through a live camera (face identity, expression, objects, gaze), keeps a memory of the relationship, and talks through a local Qwen2.5 reasoning loop. Nothing leaves the machine.

MSPI-Tiny
Distils an offline audiovisual-attention model into interpretable decision-tree students for a live camera assistant, targeting under 100 ms per frame on edge hardware.

DinoSynthesis
A phylogenetic VAE-GAN (PyTorch) that interpolates between bird and reptile vocal latents to speculate on dinosaur sounds, with its own data pipeline for collection, SNR filtering and class rebalancing.

FILMDA Player
Real-time audience measurement from blinks and facial emotion. Built for my MTech thesis on AI-driven film analytics.

• Rahulraj P. V., Sanil J., Anoop V. S., Asharaf S. Monetize the Dual: A Data Analytic Approach for Native Language and Prequel Movie Popularity Analysis. ICDAI 2023, Springer. doi
• Narrative DNA: DNA-sequence inspired analysis of narrative dynamics in movie scripts. Under review, 2026.
• JOY. MRes thesis on sequential visual attention in film, UAL, 2026.

_{Smaller experiments: GRAD0 (LLM news director) · LumiMeet · wind.ai · story-space · film_space_vr-}