Therapeutic plush bear AI

# Proposal - Therapeutic Plush AI

**Hello!**

I am Denys Hritsanov, an AI/ML specialist with 15 years of experience and the role of Head of AI Transformation at Deloitte. Your project perfectly matches my background.

## Experience with LLM
**Models I have worked with offline:**
- Llama 2/3 (7B, 13B) - fine-tuning and quantization
… - Mistral 7B, TinyLlama 1.1B, Bielik 7B
- 4-bit quantization (GPTQ), model pruning, ONNX optimization

## Voice Integration
**Complete voice pipelines:**
- STT: Whisper (offline), Wav2Vec2
- TTS: Coqui TTS, Mozilla TTS for local deployment
- Audio processing: noise reduction, VAD
- Experience with children's speech and emotion detection

## Hardware for Prototype
**Recommendation: Jetson Nano**
- 128 CUDA cores - ideal for LLM inference
- 4GB RAM sufficient for quantized 7B model
-

Show original

Translate

Good day!
My name is Valentin, and I represent Arctic Web Agency. We are a team of experienced web developers specializing in creating modern and effective web solutions for business. I can provide examples of our similar work in private messages. Ready to take on your project!

Best regards,
Arctic Web Team
Freelancehunt

Show original

Translate

Hello,
We are a company with over 10 years of experience in the market, and our team includes PhD-level experts. We have a proven track record working with LLMs, including projects we can showcase without NDA (and many more under NDA).
We can also present ready-made demos and integration examples so you can evaluate the quality and performance of our solutions. We are confident we can help bring your innovative project to life.
We would be happy to discuss the details and present a demo on the upcoming call.

Hello,
I understand that you are creating an offline voice-responsive therapeutic AI toy for children. I can help develop a local LLM pipeline running on devices such as Jetson Nano or Raspberry Pi, optimized for low latency and resource consumption.

I have worked with offline Llama and TinyLlama. For voice handling, I used Whisper and Coqui to build speech pipelines. I can also suggest emotion classifiers for detecting tone or intent from speech. For the prototype, I recommend Jetson Nano due to GPU support and flexibility in ML.

Show original

Translate

Good day, I have 5 years of experience, ready to help, please contact me to discuss the details.
I have worked with various LLM models offline, including LLaMA, Mistral, Falcon, as well as GGUF versions of models through llama.cpp, ollama, and other lightweight interfaces that allow running models on local hardware without cloud connection. I have experience optimizing models for specific hardware to achieve maximum performance without loss of response quality. I also worked with voice integration — implementing both TTS and STT, including using tools like Whisper from OpenAI, Coqui, and Vosk, which work well offline, even on low-end hardware.
For prototyping, I recommend using a Mini PC based on CPU with AVX2/AVX512 or GPU (e.g., RTX 3060–4060) — this is enough to work locally with models up to 7B in real-time, and it allows maintaining full autonomy without third-party services.
I can adapt the model to system capabilities, ensure integration with voice interfaces, build a fully offline assistant or interface that works quickly without data leaks to third parties.
Therefore, if you need a solution focused on security, efficiency, and autonomy — I am the right person to work with.

Show original

Translate

Hello!
I am interested in your project – a very cool idea, ethical and technically challenging.

Answering your questions:

1. Which LLM models do you know and have worked with offline:
I have worked with local LLMs (Quantized):
• LLaMA 2 / TinyLlama (via GGUF + llama.cpp / Ollama)
• Phi-2, Gemma, Mistral 7B, Bielik — in Q4_0, Q5_K, Q8_0 formats for running on low-power ARM/CPU devices.
… I also have experience with fine-tuning / prompt-tuning on custom datasets for specialized applications.

2. Do you have experience with voice integration:
Yes, I have created full voice pipelines:
• Whisper.cpp (local recognition)
• Silero TTS / Coqui.ai / Piper for speech synthesis
• Processing via Node.js / Python with passing to LLM via API or locally.

3. Hardware for prototype:
• Jetson Nano 4GB or Raspberry Pi 5 with active cooling.
• If low latency is important — I recommend additionally Google Coral USB for edge ML inference (emoji classification/recognition).
• Also tested Rockchip RK3588 — powerful but more complex to set up.

I would be happy to discuss technical details and potential collaboration.
I also have a background in child-safe AI and have worked with ethical frameworks (GDPR, AI Act).
Open to both contractual cooperation and participation in the project as a core developer/CTO.

Show original

Translate

Hello!

This project really caught my attention, not only because it is exciting technologically, but also because it can really make a difference for children. I can help you build a local language model that will run smoothly on compact hardware such as Raspberry Pi or Jetson Nano, and provide quick responses without excessive memory usage. I have worked with lightweight models like TinyLlama and Gemma, and I can fine-tune them to operate offline while sounding natural and warm.

Regarding the voice part, I can create a complete system that listens, understands, and responds vocally, all on the device. I would like to introduce a subtle emotional signal system that detects the child's tone and adjusts the response of the plush toy. If the child sounds uncertain or upset, the voice can soften or soothe without the need for a complex system. This adds an extra human touch, which I believe is very important for such a toy.

Jetson Nano would be my first hardware choice because it allows us to run everything locally, and if needed, it has a bit more power, especially for smoother voice playback. I can also assist in securely synchronizing everything with the parent’s app, keeping all sensitive data private on the device.

Thank you!

Show original

Translate

Hello!
I am a machine learning specialist with extensive experience working with such devices (Raspberry Pi, Jetson Nano). My main portfolio includes a project related to training models for detecting military objects, which operate locally on limited hardware.

Hardware: I recommend Jetson Nano or Raspberry Pi 5 with Coral USB, depending on the budget. Nano — efficient for TTS and STT, Coral — for accelerating inference.

Emotion/intent classification: I can suggest lightweight models derived from BERT (for example DistilBERT or TinyBERT) or rule-based fallback for critical situations.

Architecture: I propose a hybrid architecture: fully offline processing on the device + state synchronization via MQTT or BLE with the main application.

… I can provide:
Demonstration videos of target tracking from a real drone

Code examples from a real project: tracking, control, model optimization for Raspberry Pi (Python, ONNX, quantization)

Open to NDA and long-term cooperation as an AI developer / CTO

Let's create something that truly helps children.
Ready to discuss technical details, MVP requirements, and initial steps.

Show original

Translate