_____
/_____\
____[\`---'/]____
/\ #\ \_____/ /# /\
/ \# \_.---._/ #/ \
/ /|\ | | /|\ \
/___/ | | | | | | \___\
| | | | |---| | | | |
|__| \_| |_#_| |_/ |__|
//\\ <\ _//^\\_ /> //\\
\||/ |\//// \\\\/| \||/
class Harry(AIEngineer):
"""
"""
def __init__(self):
self.name = "Hari Narayan Sah"
self.bio = """
"""
self.location = "Bhaktapur, Nepal"
self.phone = "+977 9814743551"
self.languages = {
"Maithili": "native",
"Nepali": "fluent",
"Hindi": "fluent",
"English": "professional working proficiency",
}
self.education = [
{
"degree": "Data Science with Python",
"school": "Mindrisers Institute of Technology, Kathmandu",
"year": "Sep 2024 - Jan 2025",
},
{
"degree": "Diploma in Computer Engineering",
"school": "Adarsha Secondary School, Biratnagar",
"year": "2020 - 2024",
},
{
"degree": "Diploma in Graphic Designing",
"school": "New BIIT Institute, Biratnagar",
"year": "Jan 2023 - Apr 2023",
},
]
self.experience = [
{
"role": "AI Engineer",
"company": "Mantra Ideas, Lalitpur",
"year": "Sep 2025 - Present",
"notes": [
"Medical Interview AI: built the production voice interview runtime (FastAPI WebSocket, Nova Sonic speech, Claude Haiku 4.5 on Bedrock via Strands for probing, transitions, and structured answers)",
"Kept sessions usable on bad networks: quiet-moment reconnects, unspoken-speech replay, heartbeat resume, realtime transcript writes to S3",
"Shipped chat-only interviews and mid-session voice-to-chat on the same question state",
"Shipped Scribe: streaming Amazon Transcribe, speaker diarization, Haiku visit briefs, romanized transcripts",
"Built the anonymized interview corpus behind analytics and PDF reports, with PII scrub",
"Built production voice agents for a UK client (hotel orders, garage bookings) with Vapi, n8n, and Twilio",
],
},
{
"role": "Data Science Intern",
"company": "Mindrisers Technologies, Kathmandu",
"year": "Jun 2025 - Sep 2025",
"notes": [
"Generative assistants with LLaMA + LangChain: medical chatbot on the Gale Encyclopedia, news research app (Groq Llama 3 + FAISS)",
"Classical and neural models for churn prediction and financial sentiment, deployed with the engineering team",
],
},
]
self.projects = [
{
"name": "AgentBridge",
"type": "solo open-source project: one session layer for every AI coding agent",
"stack": "Rust, MCP server, PyPI (agentbridge-rs), Linux / macOS / Windows",
"highlights": "start a conversation in Claude Code and continue it in Codex, OpenCode or Antigravity, from any folder; reads sessions from the original files, writes continued turns back to the origin",
},
{
"name": "Medical Interview AI",
"type": "clinical voice interview and scribe platform",
"stack": "Nova Sonic, Claude Haiku (Bedrock / Strands), DynamoDB, S3, ECR, Bedrock AgentCore",
"highlights": "physician interviews over bidirectional audio or chat; Scribe turns a visit into a diarized, romanized brief; ARM64 image on AgentCore",
},
{
"name": "TelvoxAI",
"type": "LiveKit voice-agent platform",
"stack": "LiveKit, OpenAI, vLLM, Groq STT, ElevenLabs, Google TTS, Azure, Silero VAD, RNNoise, WebRTC NS",
"highlights": "phone agents for leads, booking, transfer, knowledge-base Q&A, and custom HTTP tools, configured per agent from an admin panel; Groq STT cut turn latency from ~8-9s to ~1-2s; filler audio covers tool-call gaps; self-hosted LLMs via vLLM",
},
{
"name": "Nepali TTS",
"type": "TTS fine-tuning for Nepali speech",
"stack": "Chatterbox multilingual (Devanagari), Matcha-TTS single/multi-speaker, Qwen3-TTS, Fish Speech",
"highlights": "models published on Hugging Face; Dockerized Chatterbox inference service with latency benchmarks and load tests",
},
{
"name": "BanknoteOCR",
"type": "Nepali denomination and signature pipeline",
"stack": "Roboflow detector, PyTorch siamese signature model, FastAPI, OpenCV, ONNX Runtime, CLIP, Chroma",
"highlights": "~90% production accuracy",
},
{
"name": "AnswerServiceAI",
"type": "UK voice workflows",
"stack": "Vapi, n8n, Twilio",
"highlights": "hotel order handling and garage booking agents with live workflow automation and handoff",
},
]
self.skills = {
"voice_agents": [
"Amazon Nova Sonic 2", "GPT Realtime", "AWS Strands", "Claude Haiku 4.5", "FastAPI WebSockets",
"Amazon Transcribe", "LiveKit", "Silero VAD", "RNNoise", "WebRTC noise suppression",
"Azure Speech", "ElevenLabs", "Groq STT", "OpenAI", "Deepgram",
],
"nepali_speech": [
"Chatterbox multilingual fine-tune", "Matcha-TTS", "Qwen3-TTS", "Fish Speech", "Devanagari romanization",
],
"llm_agents": [
"Bedrock", "LangChain", "LangGraph", "LangSmith", "LLaMA", "Groq", "FAISS", "tool calling", "MCP", "RAG",
],
"cloud_deploy": [
"AWS S3", "DynamoDB", "ECR", "Bedrock AgentCore", "Docker", "Azure",
],
"vision": [
"PyTorch", "OpenCV", "ONNX Runtime", "Roboflow", "CLIP", "Chroma", "FastAPI",
],
"voice_automation": [
"Vapi", "n8n", "Twilio",
],
"languages": [
"Python", "JavaScript", "C", "C++", "Go",
],
}
def contact(self):
return {
"phone": "+977 9814743551",
"location": "Bhaktapur, Nepal",
}
if __name__ == "__main__":
harry = Harry()
print(harry.bio)
print(harry.contact())