Vaishnav V Nair

Applied generative AI research and solution architecture for e-commerce creative production, inside Pattern Labs.

Built Pattern's foundation for label-accurate product generation. Promoted to Principal in about six months.

AI creative studio. Artistic QR codes, AI product photography, animation. Founded at 22.

Built the model-training and prompt workflows behind daily news, social and ad content for the Stance app.

End-to-end image-stack generation. One brand input becomes the complete 7 to 9 image listing stack, at bulk scale. art director agent · deterministic render · multi-agent pipeline

▶Explainer film

• An Art Director agent designs every image slot from scratch against 500+ design rules, produces three genuinely different variations, then screenshots and vision-scores its own work until it clears the quality gate.
• A strict Auditor agent scores every image on nine criteria and rules SHIP, ITERATE or REDESIGN. Supporting agents feed the flow: strategy, iconography, image generation and template supply.
• Deterministic where it matters: visual hierarchy, legal copy and text fit are guaranteed by construction. Creative scenes come from image models. Text is real HTML, never AI-rasterized.
• Measured on a live ten-ASIN brand: $651.62 down to $136.58, and 24.1 hours down to 6.1. Per ASIN, that is $65 to $14 and 2.4 hours to 37 minutes. Actual production numbers, not estimates.

A synthetic focus group for creative work. agents · personas · evaluation

▶Explainer film

• Every agent holds a distinct customer or expert persona and reviews the creative from that perspective. The swarm ranks variants and flags design issues before a human ever looks.
• The hard parts were never the agents. Personas are engineered to disagree usefully instead of converging on one bland verdict, and the comparison structure keeps ranking stable rather than order-dependent.
• Self-initiated. Once generation stopped being the bottleneck, review became the bottleneck, and nobody had noticed yet.
• Next milestone: validating its predictions against real human preference data.

The high-frequency detailer. Two models doing one job each, 100% label accuracy in internal review. lora · segmentation · qwen · patent filed

▶Explainer film

• Diffusion models degrade exactly the detail that carries brand identity: packaging text, logos, icons. The highest-frequency information on the object, and the only part a brand's legal team will look at.
• Stage one generates with a LoRA trained on the product category. Stage two is a Qwen-based restoration pass: segmentation finds only the printed regions and matches them back against the reference.
• Generalises across bottles, sachets, jars, clothes, bags and tools, and became the foundation Pattern's production generation is built on.
• Patent application filed for the high-frequency detail restoration workflow. Named inventor.

I changed the camera, not the prompt. The capture specification behind Pattern's robotic photography rig. capture · lora · hardware

▶Explainer film

• The generations kept failing in specific, repeatable ways, and the instinct in the room was to write better prompts. The failures were coverage failures: wrong angles, wrong lighting, not enough frames of the parts that mattered.
• The spec defines lighting configuration, camera angles, capture distances, close-up versus wide coverage, and how diverse the photo set must be, all derived from LoRA training experiments.
• It informed the design of Portal Studio, the robotic capture rig. Hardware decisions made on model evidence.
• The Portal won Pattern's Innovation Cup 2025 and was demoed live at Accelerate 26, including a generation run for the CEO.

Portal Studio ↗Showcase film ↗

The va1 node pack, public on the Comfy Registry. 1.3K downloads. comfyui · python · open source

▶Explainer film

• Pad Image by Aspect for Outpaint: expands, pads and masks images to fixed or randomized aspect ratios, with precise spatial and scale control for outpainting and canvas expansion.
• Image Mask Comparer: compares the masked regions of two images with similarity scoring, and automatically retries or re-queues on mismatch.
• Behind the public pack sit production workflows for relighting, background replacement and image expansion across Flux, Qwen Image Edit and Nano Banana.

GitHub ↗Comfy Registry ↗

An AI creative studio, founded at 22. Product photography, animation, model training, and QR art at national-campaign scale. founder · studio · controlnet

▶Explainer film

• Model training and fine-tuning were the core service, across product photography, animation and QR art.
• A ControlNet QR workflow of my own tuning turned functional codes into artwork clients commissioned. Top-rated QR artist on the platform.
• Tata IPL 2024, for TATA NEU: 20+ QR codes rolled out across tier-one cities, online and in stores including Westside, 1mg, Croma and BigBasket.
• Myntra bulk lifestyle imagery: one workflow generating AI models wearing the catalogue across dress categories.

codeartx.com ↗Fiverr reviews ↗Instagram ↗

A personal AI video pipeline. Script to finished short-form film in one run. video · renderer · tts · personal build

▶Explainer film

• Script, voice, captions and motion produced by one deterministic pipeline: a code-driven renderer, local text-to-speech, word-timed captions and clip retrieval, composed per content profile.
• Pacing, typography and motion are measured from reference films and enforced as constraints, not eyeballed.
• Built to publish without a human in the loop, and to render the explainer films on this page.