Lily Software Inc

Turn complex models into business outcomes

Lily Software Inc helps teams design, build, and ship software products and machine learning systems—from music websites and mobile apps to research-grade LLMs—with clarity and measurable impact.

About

About Lily Software Inc

Lily Software Inc is a U.S. software company based in Eastvale, California. We develop music-related websites and mobile applications, and we deliver applied AI and large language model (LLM) software across data science, natural language processing, speech, and computer vision.

Our teams work end to end—from product design and engineering through model training, evaluation, and production deployment—so clients can ship reliable digital products and AI features with measurable business outcomes.

Our official website is www.lilysoftware.us.

Services

What we deliver

Lily Software Inc builds music digital products and applied AI systems. Explore our service lines below, then dive into our detailed practice areas for data science, NLP, speech, and image processing.

  • Music Web & Mobile Apps

    We design and develop music-oriented websites and mobile applications—content and discovery experiences, companion backends, admin tools, and iOS / Android clients that stay in sync with your platform.

    • Music websites & portals
    • iOS & Android apps
    • Streaming & community features
    • Content & admin backends
  • LLM & Applied AI

    We help organizations apply large language models and related AI in real products: model integration, domain adaptation, evaluation pipelines, and multimodal features spanning language, speech, and vision.

    • LLM integration & fine-tuning
    • Evaluation & alignment
    • Conversational & content AI
    • Multimodal systems
  • Custom Software Development

    Full-stack web and backend systems, cloud deployment, product iteration, and technical consulting—from early prototypes through production maintenance.

    • Web & backend systems
    • Cloud deployment
    • Product iteration
    • Technical consulting

Practice deep-dives: Data Science Natural Language Processing Speech Processing Image Processing

Analytics dashboards and data visualization representing enterprise data science and platform work
01

Data Science

We develop and optimize domain-specific large language models (LLMs) and small language models (SLMs) to address unique industry challenges and deliver practical, innovative solutions. Our work spans building advanced SLMs and LLMs for targeted domains, with emphasis on supervised fine-tuning, reinforcement learning from human feedback (RLHF), pre- and post-training, rigorous evaluation, and multi-modal integration where the problem demands it.

  • Domain-specific LLMs & SLMs
  • Supervised fine-tuning
  • RLHF
  • Pre- & post-training
  • Evaluation & alignment
  • Multi-modal pipelines
  • LLM pretrain & fine-tune (edge-scale)

    Built an 8-layer transformer with Mixture-of-Experts at roughly 145M parameters, pretrained primarily on Chinese text, then fine-tuned on a Qwen distillation dataset with RLHF. The resulting model delivers GPT-like capability with reduced size and training cost, suited to edge deployment.

  • John Deere — big data & ERP modernization

    Built a big data platform and upgraded a legacy ERP stack: designed a data warehouse, data cubes, and a data visualization portal, and moved the ERP footprint to a .NET-based platform for long-term maintainability.

  • Citibank — iOS wealth management platform

    Delivered an iOS-based wealth management experience spanning UI, SOA, and mobile engineering. The system supports privileged clients with stock and options trading and purchasing bank wealth-management products.

Notebook and writing tools suggesting language, dialogue, and NLP systems
02

Natural Language Processing

End-to-end language AI for customer service, marketing, and research: intent recognition, retrieval and matching, empathetic dialogue, and efficient adaptation of large models—from BERT-era encoders to modern LLMs—with attention to latency, accuracy, and measurable business metrics.

  • Agentic AI
  • Deep question answering
  • Prompt tuning & optimization
  • Pretrain & fine-tune LLMs
  • LoRA & PEFT
  • RLHF
  • Multimodal text + graph
  • Intelligent customer service (SLM) — Konka

    Integrated an intelligent service stack into the client’s EIP for intention recognition and accurate problem–solution matching. Used call-center data to train BERT for query matching. Deployed interception rate above 90% in Konka’s call-center environment.

  • IBM intelligent customer service

    Built a system that partitions knowledge into hardware, software, servers, and services so each human–machine turn moves the dialog closer to the right knowledge artifact.

  • Pfizer — multilingual product marketing

    Delivered a multilingual product-marketing platform with a content-based management model, supporting a distributed team across India, Europe, China, and the USA.

  • Emotion classification (multi-modal & graph)

    Engineered a multi-modal emotion stack combining CRNN, DNN, Transformer, and XGBoost for verbal and non-verbal cues. Implemented a prompt-based PyTorch pipeline using BERT (text), ResNet (image), and VGGish (audio), augmented with graph neural networks to model inter-modal relationships on standard emotion benchmarks.

  • Empathetic response generation

    Developed an LLM-augmented method for empathetic replies in knowledge-grounded dialog. Combined COMET commonsense graphs and persona attributes in rich prompts to synthesize training data, then trained a variational autoencoder (VAE) in PyTorch on the enhanced set for stronger empathy and downstream metrics.

  • Llama 3 — LoRA & full fine-tuning

    Applied LoRA to Llama 3-8B and full fine-tuning on Llama 3-3B using 52K Alpaca-style instruction samples for instruction-following quality.

Studio microphone and audio equipment representing speech and voice AI
03

Speech Processing

Wake-word and speaker verification for consumer devices, speech enhancement for edge deployment, and scalable ASR with real-time hotword and moderation hooks—from data collection at scale to containerized inference.

  • Wake word & KWS
  • Speaker verification
  • Multi-task & cross-attention
  • Speech enhancement
  • Lightweight Transformers
  • ASR & Kubernetes
  • Wake word & speaker verification (SLM)

    Built a deep cross-attention architecture with multi-task learning to improve robustness for edge keyword spotting and speaker verification together. Collected 240K wake-word utterances from four Konka-branded TVs, then trained a joint ConvMixer + ECAPA-TDNN stack for stronger dual-task performance.

  • Speech enhancement (SLM)

    Developed a lightweight multi-axis Transformer with hybrid attention for time–frequency modeling, cutting compute for edge speech enhancement while preserving quality.

    Lite-RTSE project details →

  • ASR system (SLM)

    Stood up scalable ASR on Kubernetes with containerized WeNet models for real-time hotword detection and content moderation, plus a web-based admin UI for client configuration. Improved dialect recognition by training on roughly 600 hours of region-specific speech.

Abstract visualization suggesting computer vision, generative models, and multimodal AI
04

Image Processing

Vision–language models (VLMs), diffusion fine-tuning, and research-grade image generation stacks—from parameter-efficient multimodal training under tight compute budgets to GRPO and video understanding for safety-critical analytics.

  • VLM pretrain & SFT
  • CLIP-style fusion
  • Diffusion & Dreambooth LoRA
  • Image & video understanding
  • Token-based image generation
  • VLM pretrain & fine-tune (<50M params)

    Built a parameter-efficient multimodal (vision–language) model under 50M parameters: frozen CLIP visual encoder, transformer language backbone, special-token fusion, and trainable projection layers. Stage 1: pretraining on ~570K images from CC-3M and COCO 2014 with frozen backbone. Stage 2: supervised fine-tuning on ~300K instruction pairs and ~150K images from llava-en-zh-300k with full-parameter optimization.

  • FLUX & custom concepts

    Fine-tuned the FLUX 12B diffusion model with Dreambooth LoRA on 4–8 images per custom visual concept, wired ComfyUI for inference, and shipped 1024×1024 concept renders for production-style workflows.

  • Qwen2.5-VL-7B & InternVL3-8B

    Wildfire detection with Qwen2.5-VL-7B and GRPO on ~42K wildfire-prediction images. Dangerous-driving analysis with InternVL3-8B on ~1.5K video samples from a Nexar collision-prediction corpus.

  • Minimal DALL-E & CLIP research

    Implemented a minimal DALL-E–style pipeline for efficient inference: BART encoder for text embeddings, BART decoder to predict image tokens, and a pretrained VQGAN decoder for high-resolution reconstruction. Additional CLIP training and inference research.

For image search, retrieval, and catalog-scale embedding strategies, we apply the same embedding and evaluation rigor used in our VLM and multimodal engagements.

Contact Lily Software Inc

Tell us about your goals, timeline, and constraints. We’ll respond with a clear path forward—whether that’s a short advisory sprint or a longer build.

Email xingwei@lilysoftware.us