Job Specifications
Embedded Generative AI Developer - Internship (Generative AI Photo)
About Lillup
Lillup is redefining human capital management with embedded, on-device AI. We build intelligent, scalable systems that run privately on smartphones, computers, and constrained hardware. As an intern, you'll work on real product challenges in a fast, collaborative environment alongside engineers and designers in AI, blockchain, and UX.
Role Overview
We're looking for a motivated Embedded Generative AI Developer focused on Generative AI for Photos--from on-device image generation to editing, enhancement, and privacy-preserving transformations. Your work will directly power the Talent Passport across iOS, Android, and Web with low-latency, real-time experiences.
Primary LLM (for orchestration/agent flow): Google Gemma (on-device where possible).
Scope: Generative Photo (prompt-to-image, image-to-image, background/subject edits, style transfer, anonymization).
Key Responsibilities
Build Embedded Generative Photo Pipelines
Design, implement, and optimize on-device photo generation/editing (diffusion-based pipelines, img2img/inpainting, upscaling).
Integrate Gemma as the local controller/agent for prompt parsing, tool routing, and safety checks.
Optimize for Mobile
Achieve sub-second UX where feasible via quantization (INT8/mixed FP16),etc.
Leverage hardware acceleration: iOS (Core ML + Metal/MPS), Android (NNAPI/Vulkan/TFLite GPU), Web (WebGPU/WASM).
Product Integration
Collaborate with AI engineers and UX to ship photo flows: prompt-generate/edit-preview-commit, with safe defaults and rollbacks.
Implement privacy features and on-device safety filters.
Testing & Continuous Tuning
Create device matrices and latency/quality benchmarks (iPhone/Pixel/mid-tier Android).
Build real-time feedback loops for quality scoring (CLIP-like similarity, perceptual metrics) and prompt refinements via Gemma.
Documentation & Research
Document model variants, quantization recipes, schedulers, and acceleration tricks.
Qualifications
Solid understanding of Generative AI and diffusion models (img2img, inpainting, SR, style transfer).
Hands-on with on-device inference (Core ML, TFLite GPU, WebGPU) and model compression/quantization.
Comfortable wiring an LLM controller (Google Gemma) to toolchains and safety layers.
Experience with mobile or web graphics pipelines (Metal/MPS, Vulkan, WebGPU) and efficient memory management.
Strong autonomy, curiosity, and shipping mindset in a remote, fast-paced setup.
Nice to Have
Experience with LoRA fine-tuning, ControlNet/IP-Adapter, prompt schedulers, or custom UNet/VAE graph conversions (Core ML / TFLite / ONNX).
Image quality evaluation (LPIPS, FID proxies, CLIPScore) and prompt-programming for visual tasks.
Basic VLM familiarity (e.g., using a VLM for auto-captioning, safety checks, or prompt rewrites).
What We Offer
Real impact on a shipped, on-device product.
Remote-first, flexible schedule with a global team.
Mentorship from engineers working at the edge of embedded AI.
Note: This is an unpaid, training-focused internship. Exceptional performance may lead to future opportunities.
How to Apply
Send your resume and a brief cover letter that includes:
Your experience with on-device generative photo workflows.
How you've used Google Gemma (or how you'd wire it as an on-device orchestrator for photo tasks).
Links to relevant projects (GitHub, TestFlight/APK builds, demos) showing:
Core ML/TFLite/WebGPU conversions and speedups
Before/after examples (inpainting, SR, background edits)
Any LoRA/ControlNet/IP-Adapter work
Apply now--let's build the future of embedded, privacy-first generative photo together.