Selected work
Production AI systems and research builds across 2PointZero, MBZUAI, VisionLabs, the University of Peradeniya, and independent work.
2PointZero needed a way for an "AI delegate" to engage in real time with C-suite from 20+ portfolio companies during the Growth Day event - a setting where latency, faithfulness, and an embodied presence all mattered more than batch chat quality.
Workforce benchmarking across 2PointZero's portfolio companies was a multi-week manual exercise - and the source data was sparse and unevenly disclosed across companies. Stakeholders needed a conversational way to ask FTE questions and trust the answers.
Portfolio analytics agents at 2PointZero were being tested ad-hoc - different prompts, different runs, no shared ground-truth, and no way to compare providers or regression-test prompt changes.
Legal requests at 2PointZero arrived as ad-hoc emails. No structured fields, no SLA visibility, no consistent place to drop attachments. The legal team spent more time triaging than reviewing.
University career-fair recruiters scan dozens of resumes per booth in minutes, with QR codes pointing to a mess of formats (raw PDFs, Drive links, Symplicity redirects, GitHub, LinkedIn, portfolio sites). There is no time to read each one, and model latency would block the next candidate if analysis ran inline.
IHC's agentic AI platform (SAIF) needed thousands of custom GPTs configured per role and function across portfolio companies - a months-long manual configuration job at human scale. Each portfolio company had its own sectoral nuance (retail, beauty & wellness, media, etc.) that off-the-shelf assistants did not capture.
Evaluations of large multimodal models (LMMs) report a single accuracy number per benchmark, but those numbers are silently a function of the prompt phrasing. A robust LMM should not swing 15 points on the same MCQA question because of cosmetic prompt edits - and the field had no systematic way to measure how much it actually does.
An 8-brand, 134-branch UAE salon group needed objective evidence about which branches to retain, consolidate, or relocate, and where to expand. Existing decisions were anecdotal; financial signals alone missed location dynamics.
Existing Video-LMM benchmarks focused on short clips (seconds to a minute), but downstream applications - egocentric assistance, surveillance, instructional video - require reasoning over 3-30 minute footage. There was no systematic way to evaluate long-form video understanding, and prior benchmarks under-sampled egocentric content entirely.
ConvNeXt-V2 is competitive on ImageNet but loses ground on fine-grained classification (CUB Birds, FGVC Aircraft, FoodX) where class-discriminative features live in narrow channel subspaces. Adding generic attention helps a little but tends to inflate FLOPs without proportionate accuracy gain.
Face-recognition models trained on real-only data plateau at the long tail of identity - small numbers of images per identity, demographic skew, and limited pose/lighting coverage. Adding synthetic identities is appealing, but only works if the synthetic generator is conditioned on a representative descriptor of each real identity.
Surveillance video in the wild has multiple two-person interactions happening simultaneously in a crowded frame - and most action-recognition pipelines were trained on single-actor clips, so they collapse when interactions overlap. A real-time pipeline needed to localize each pair, isolate them, and classify the interaction without dropping frames.
Video-ChatGPT was designed around short clips. On long videos (multi-minute), token budgets exploded and frame-level attention degraded; meanwhile the model had no audio channel, so dialogue and ambient sound that often disambiguate a scene were invisible to it.
Solar irradiance (GHI) varies sharply with sky condition - clear, cumulus, overcast - and a single regression network has to learn three different physics regimes at once, which it does poorly. Photovoltaic operators need short-horizon forecasts that are accurate across all three regimes, not just on average.