Quality after the demo
The hard part of AI product work is deciding what good means, measuring it, and making the system useful in a real workflow.
AI Engineer
Gaurav Pooniwala
I build AI products and the systems around them. I am currently a founding engineer at Cubic, working on AI code review. Before that, I helped build generative video systems at Rephrase AI before its acquisition by Adobe, worked on deep learning systems at Samsung R&D, and built AI products in early-stage startup environments.
Now
Building at Cubic
Founding engineer leading AI development for code review.
Range
Models to products
Model behavior, evaluation, product quality, user experience, and business constraints.
Before
Rephrase AI -> Adobe
Built generative video systems before the company was acquired by Adobe.
Lens
Engineering + business
Deep learning research, startup execution, leadership, and business strategy.
Now
Building at Cubic.
Code review is the current problem. The broader work is AI product engineering: model behavior, evaluation, cost, product quality, and developer experience.
AI code review for complex codebases
Cubic builds AI code review for teams working in complex repositories. It has been publicly ranked #1 on an independent AI code review benchmark. I lead AI development across the places where model behavior turns into product quality.
Models, product, and business constraints
The pattern across my work is taking AI from a promising behavior to a usable system: what to measure, what to ship, what to simplify, and what needs to be cheaper or more reliable.
Speaking and writing
Current themes.
Most AI conversations get stuck at capability. I am more interested in the work after that: evaluation, reliability, cost, product judgment, and how teams change when AI becomes part of the workflow.
Evaluation that changes decisions
Promising model behavior only matters after it survives cost, latency, edge cases, trust, and repeated use.
AI inside engineering teams
When generation gets cheaper, judgment, review, ownership, and product taste become more important.
Track record
Proof points.
Cubic
Leading AI development for code review, with work spanning model behavior, evaluation, cost, and developer experience.
Cost and reliability
Reduced AI system running costs by 3x while keeping attention on product quality and reliability.
Rephrase AI -> Adobe
Helped build generative video and lip-sync systems before the company was acquired by Adobe and its technology became part of the Firefly story.
Nexcade and KIRO
Built agentic quotation workflows, RAG systems, and production AI infrastructure in early-stage environments.
Samsung R&D and IIT Bombay
Started in deep learning research and optimization, with a technical foundation shaped by Samsung R&D and IIT Bombay.
Start a conversation
For talks, collaborations, advisory conversations, or selective senior opportunities, send a short note with the audience and context.