Amar S.

I build production-grade Multimodal AI pipelines - systems where LLMs, image generation, video, and audio work together as one reliable product, with evals built in so quality is a measured number. Photo in, personalized story and consistent illustrations out. Most AI projects stall at 80%. The prototype works, then it never becomes a product clients can trust: characters drift between images, outputs break schemas, accuracy is a guess. That last stretch - consistency, structured outputs, evaluation, real deployment - is exactly what I do. 10+ years in software, now focused entirely on Multimodal and Generative AI. 𝐖𝐡𝐚𝐭 𝐈 𝐁𝐮𝐢𝐥𝐝: - Multimodal AI pipelines: end-to-end systems combining LLM story/content generation (structured JSON schemas) with image generation - identity-locked characters across hundreds of pages, outfit consistency, emotion and expression control, multi-character scenes, reference-image workflows. - AI image generation pipelines: ComfyUI, Flux / FLUX.2, SDXL, Nano Banana - with LoRA training (Kohya, ai-toolkit), ControlNet, IPAdapter, PuLID/InstantID face locking, inpainting, and Runware/API-based generation for production consistency and quality control. - AI video generation pipelines: image-to-video and text-to-video using WAN 2.2, LTX, Kling, Veo, Runway, Seedance - plus audio-driven lip-sync, talking avatars, and motion-controlled animation systems. - AI UGC and ad creatives: scroll-stopping product showcase videos, UGC-style ads, and short-form marketing content for TikTok, Reels, and Meta campaigns - built as repeatable pipelines. - Audio & speech: Whisper speech-to-text, ElevenLabs TTS and voice cloning, photo & audio to talking-avatar video pipelines. - LLM & VLM integration: GPT-4o/5-class vision, Claude, Gemini integrated into business systems - structured outputs, schema validation, RAG with LangChain/LangGraph and vector DBs (Pinecone, FAISS). - LLM/VLM Evals & quality optimization: golden datasets, LLM-as-judge and VLM-as-judge scoring, experiment-driven prompt optimization, regression testing with Langfuse/Promptfoo/Braintrust - so accuracy is tracked as a number that moves ("71% to 84% in three cycles"). - Full-stack development: React/Next.js frontends, Node.js/FastAPI backends, PostgreSQL - so your AI product ships as one working system. - Deployment & MLOps: FastAPI, Docker, GPU infrastructure and queue serialization, cost-efficient serverless GPU inference on RunPod/Modal, and CI/CD for ML pipelines. 𝐑𝐞𝐜𝐞𝐧𝐭 𝐖𝐨𝐫𝐤: ✔️ 𝗣𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘇𝗲𝗱 𝗔𝗜 𝗦𝘁𝗼𝗿𝘆𝗯𝗼𝗼𝗸 𝗣𝗹𝗮𝘁𝗳𝗼𝗿𝗺: children's book illustration pipeline - hero photo & child's details in; stylized character reference, structured 15-page story JSON, and identity-consistent page illustrations out (Flux & Runware). Personalized storybooks with character, outfit, and expression consistency at production scale. ✔️ 𝗗𝗔𝗧𝗔𝗜𝗦𝗠: AI character generation platform with dual ComfyUI pipelines, automated LoRA training on RunPod GPUs, quality gating, and GPU queue serialization - thousands of identity-locked images. ✔️ 𝗔𝗜 𝗜𝗻𝗳𝗹𝘂𝗲𝗻𝗰𝗲𝗿 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗼𝗿: generative AI layer for a SaaS platform - 7+ foundation models orchestrated with prompt-based auto-routing, reference-conditioned character consistency, plus motion-transfer and lip-sync pipelines. ✔️ 𝗩𝗟𝗠 𝗔𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻: eval pipeline for a food-photo-to-calories feature - golden dataset, automated scoring, experiment-driven prompt iteration to lift real accuracy on held-out data. ✔️ 𝗝𝗼𝘆𝗣𝗶𝘅: AI talking-video platform - photo & audio in, lip-synced avatar video out, identity preserved at scale. ✔️ 𝗛𝗶𝗿𝗲𝗛𝗮𝗿𝗿𝘆: multi-agent AI platform backend - LLM task automation with production-safe structured outputs and deterministic state transitions. You'll get daily updates, clear milestones, and a fully documented handover - so your team can run the system without me. If your project involves multimodal AI, image or video generation pipelines, LLM/VLM integration, evals, or getting any of it actually deployed and reliable - send me the details and I'll tell you exactly how I'd build it. 𝐓𝐞𝐜𝐡: ComfyUI, Flux, FLUX.2, SDXL, Nano Banana, Stable Diffusion, LoRA, Kohya, ai-toolkit, ControlNet, IPAdapter, PuLID, InstantID, Runware, WAN 2.2, LTX, Kling, Veo, Runway, Seedance, Whisper, ElevenLabs, FFmpeg, PyTorch, GPT-4o, Claude, Gemini, Llama, LangChain, LangGraph, RAG, Pinecone, FAISS, MLOps, FastAPI, Python, Node.js, React, Next.js, PostgreSQL, Docker, GPU, RunPod, Modal, AWS. Note: I don't take on NSFW, adult, or explicit content projects.