S">S"> Sivakumar Portfolio
Sivakumar Portfolio
Enterprise GenAI · Design Platform · 0→1 · Global · Leap 100

Ability to Enterprise
GenAI Design

First GenAI platform formally adopted in automotive manufacturing. Stable Diffusion, Flux, LoRA fine-tuning per brand, ComfyUI migration for complex workflows, multi-model handling. Included in Renault's Leap 100 programme to manufacturing a vehicle in under 100 weeks vs the historical 3.5 years.

Stable Diffusion · Flux ComfyUI · Multi-workflow LoRA per brand RabbitMQ · GCP VM Leap 100 · 5 countries · GDPR
Designers
250+ · 5 countries
Designs Generated
500,000+
Stack
Stable Diffusion · Flux · ComfyUI · A1111 · RabbitMQ · GCP
Outcomes
500K+
designs generated
all studios
−70%
design-to-concept
cycle time
5.8
avg iterations to success
from 8 baseline
−2.2 iterations
4.6
CSAT
from 3.8 at launch
−36%
GPU cost per output
₹210 → ₹135
North Star Metric
Iterations to Successful Design Output
Not output volume (incentivises noise) and not render speed (optimises for the wrong thing). The only metric that captures creative usefulness: how many attempts before a designer has something they'd present to a design lead? Baseline: 8. Current: 5.8. Target: 4 to 4.5.
User Personas
Senior Designer · Paris · 15+ yrs

Mathieu

Pain: "AI won't draw cars to that's our job." Identity threat. Initial resistance.

After 6 months: Requesting expanded access daily. Internal champion.

Design Lead · India · 8 yrs

Priya

Pain: Mon to Thu on initial concept sketches alone. 5 different tools in current workflow.

Needs: Rapid generation. Single unified tool. Brand-consistent without manual injection.

UX Automotive Designer · Global · 7 yrs

Jordan

Pain: Cross-studio brand consistency is manual. Each studio has its own style conventions.

Needs: Studio-level LoRA conditioning. Creative control with brand guardrails.

Platform Screens
Ability to design platform home screen

Ability to design platform home screen

Design generation interface
ComfyUI workflow graph
At the Leap 100 Event
Sivakumar at Leap 100 AI COE event

Presenting Ability as part of the Leap 100 AI Centre of Excellence programme

Multi-Model Architecture

Why we migrated from A1111 to ComfyUI: A1111 is excellent for standard text-to-image but hits ceilings with complex multi-step workflows to combining sketch-to-image, inpainting, style conditioning, and upscaling in a single run. ComfyUI's node-based graph allows any model combination to be chained visually by designers themselves, without engineering involvement. Result: designers build their own pipelines for specific use cases (exterior sketch → rendered concept → brand variant) in hours, not days.

Design Interface · ComfyUI Node Graph
Visual workflow builderDesigner-built pipelines · no codeSketch-to-design · Inpainting · BlendStyle reference injection
↓
Multi-Model Layer
Stable Diffusion XL · base generation Flux · high-fidelity detail rendering LoRA per brand · Renault / Nissan / Alpine ControlNet · sketch-to-image guidance CLIP interrogator · reverse prompt engineering Studio-level style conditioning
↓
Request Queue · RabbitMQ
Async job queue · priority by seniorityGCP VM instance · GPU pool managementEstimated wait time display
↓
Enterprise Compliance · GCP + On-premise
Training on Renault private cloud only (on-premise) Inference on GCP VM instances Per-studio access control via RBAC GDPR asset pipeline · AES-256 encryption Full IP audit trail · data lineage
Architecture Decisions
DecisionChosenRejectedRationale
Training data locationRenault private cloud onlyGCP public trainingDesign assets are trade secrets. Legal required training to stay on-premise. Inference on GCP only.
Workflow toolComfyUI node graphA1111 onlyA1111 ceiling for complex multi-step workflows. ComfyUI enables designer-built pipelines.
Style conditioningPer-brand LoRA fine-tuningGeneral SDXL onlyGeneric models produce generic automotive design. LoRA per brand maintains brand vocabulary.
Queue architectureRabbitMQ async job queueSynchronous requestsGPU rendering is slow. Async queue with wait-time display prevents UI blocking and sets expectations.