← Portfolio
Startup · Boat4All × ITDEV · 2024–2025
AI Companion for Kids —
Screen-Free Learning Device
AI-powered children's teddy bear — advisor, friend and tutor for the child, away from screens and phones.
🤖 AI Engineer
🫐 Raspberry Pi
🗣️ TTS / STT
🔒 Content Safety
⚡ Latency Optimization
🧠 LLM Research
100%
on-device
AI pipeline
Σύνοψη — για recruiters
AI Engineer · Startup · Boat4All × ITDEV · 2024–2025
I took full ownership of the AI layer for an embedded children's product: selected the LLM after extensive market research, designed the Python architecture (TTS/STT), integrated a content-safety layer, and optimized for Raspberry Pi.
Αντίκτυπος
Μέσα σε 1 μήνα: αρχιτεκτονική AI pipeline, LLM evaluation, TTS/STT σε Raspberry Pi, content safety module & latency optimization — χωρίς προηγούμενο AI expertise στην ομάδα.
LLM Evaluation & Selection
Embedded AI (Raspberry Pi)
TTS / STT Pipeline
Content Safety Engineering
Latency & Memory Optimization
Python Architecture
🏢 Το Startup & η Ομάδα
Who was behind the project
Η ιδέα γεννήθηκε από τον ιδρυτή της
Boat4All σε συνεργασία με την
ITDEV — ένα παιδικό αρκουδάκι με ενσωματωμένο AI που θα μιλούσε με το παιδί, θα το συμβούλευε, θα του έκανε παρέα και θα το βοηθούσε στα μαθήματα,
μακριά από κινητό και οθόνη.
Κύριος Επενδυτής
Ιδρυτής Boat4All
Project initiative and funding
Project Management
Ιδιοκτήτης ITDEV
Business direction and team coordination
Software Engineer
Senior SWE (τράπεζα εξωτερικού)
Backend development — no AI expertise
Hardware Engineer
Senior Μηχανικός ΕΑΣ
Embedded systems and hardware (Hellenic Defence Systems)
AI Engineer ★
Σπήλιος Δημακόπουλος
Responsible for the entire AI software — the only team member with AI expertise
AI Engineer (junior)
Φοιτητής Εφ. Μαθηματικών
Participating in the AI section for hands-on experience
I was involved from the first meeting to plan development. I was the only member with AI expertise — the senior software engineer had no AI knowledge, so the entire AI layer was exclusively my responsibility. Collaboration was mostly remote, with regular meetings at ITDEV's offices.
💡 Το Προϊόν
What we set out to build
A children's stuffed teddy bear with embedded AI — no tablet, no phone, no screen. A physical object the child can hug and talk to naturally.
Natural Conversation
The child speaks, the bear listens and responds in real time via on-device TTS/STT pipeline.
Homework Help
Answers questions, explains concepts, teaches — adapted to the child's age.
Companionship & Advice
Companionship, emotional support and guidance in the child's everyday situations.
Content Safety
Strict content filtering — no inappropriate content, full parental control.
Screen-Free
Digital interaction without screens — the core product philosophy.
Offline-First
On-device AI without cloud dependency — works even without internet.
⚙️ Το AI Pipeline
End-to-end architecture
💡 In plain words: STT = converts speech to text, TTS = converts text to speech. An LLM (Large Language Model) is the AI model that generates the replies; GPT is one well-known family of such models (by OpenAI).
Voice Pipeline · On-Device · Raspberry Pi
01
Audio Input
Microphone
Captures the child's speech in real time.
02
STT
Whisper
Converts speech to text, on-device.
03
Safety Filter
Blocklist
Screens content before it reaches the LLM.
04
LLM
Quantized
Generates the reply — a quantized model for a low memory footprint.
05
TTS
Piper / Coqui
Converts text into natural-sounding speech.
06
Audio Output
Speaker
Plays the reply back through the speaker.
🛠️ Τι Έκανα ως AI Engineer
Technical contributions over ~1 month
LLM Market Research
Research
Evaluation
Researched the entire LLM market (open-source & commercial) for the best fit: low latency, small memory footprint, offline capability, child safety. Evaluated Mistral, Phi, Gemma, TinyLlama and others.
Raspberry Pi Selection
Hardware
Benchmarking
Compared different Raspberry Pi models as the embedded compute medium — RAM, CPU, temperature, power consumption, cost and compatibility with AI workloads.
Αρχιτεκτονική Κώδικα
Python
Modular Design
Led the design of the AI pipeline architecture in Python — modular structure for easy component replacement (LLM, TTS engine, STT engine).
TTS / STT Development
Whisper
Piper
Coqui TTS
Developed and tested text-to-speech and speech-to-text pipelines in Python — evaluated Whisper, Piper, Coqui TTS and others for on-device use without cloud dependency.
Content Safety Layer
Blocklist
Custom Filter
Safety code with a blocklist of words and phrases — small LLMs require tight control as they lack the guardrails of larger models. Custom filter for child-safe content.
Latency Minimization
Quantization
Streaming
Buffer Mgmt
Optimized the entire pipeline to reduce end-to-end latency (STT → LLM → TTS) — quantization, streaming responses, buffer management.
Raspberry Pi Testing
Benchmarking
Stress Tests
Tests on real hardware: benchmark inference speed, memory profiling, thermal throttling analysis, and stress tests to verify on-device AI viability.
Upon my departure, the core on-device code was nearly complete — only further performance optimizations remained. Meanwhile, the team had already begun planning a second mode: a cloud-based version with API integration for stronger responses and scalability.
🧠 LLM Evaluation
Model comparison for embedded children's product
Evaluation based on criteria critical for this specific use case: on-device inference, low memory, low latency and child-appropriate language.
| Model |
On-Device |
RAM Footprint |
Latency RPi |
Child Safety |
| TinyLlama 1.1B |
✓ |
~900 MB |
Χαμηλό |
~ Αδύνατο |
| Phi-2 / Phi-3 Mini selected |
✓ |
~1.5 GB |
Μέτριο |
✓ Καλό |
| Gemma 2B |
✓ |
~1.8 GB |
Μέτριο |
✓ Καλό |
| Mistral 7B (quant) |
~ |
~4 GB |
Υψηλό |
✓ Εξαιρετικό |
| GPT-4o (cloud) |
✗ |
API |
Network-bound |
✓ Εξαιρετικό |
* Δεδομένα κατά προσέγγιση βάσει δοκιμών — η τελική επιλογή έγινε μετά από εκτεταμένο testing στο hardware-target
In just one month, as the only AI specialist on the team, I laid the foundations for an on-device AI children's product — from LLM evaluation and Raspberry Pi testing to content safety and latency optimization. The core on-device component was nearly complete upon my departure.
Spilios Dimakopoulos · AI Engineer — Startup Boat4All × ITDEV · 2024–2025