Shishir Poudel

ML Engineer · Real-time voice AI · LLMs & RAG · Speech models

[email protected]shishir-poudel.com.np github.com/Shishir3D linkedin.com/in/shishir3dNepal

Open to ML engineering and speech AI roles.

1,000+concurrent users on the voice AI I own
500,000+downloads of the app it powers
< 700 msvoice-to-voice target per turn
5AI and product systems across two roles

Summary

ML engineer who takes AI from prototype to production. I own a real-time voice-to-voice AI system (speech-to-text → LLM → text-to-speech) that serves 1,000+ concurrent users on self-hosted infrastructure in Nepal, integrated into a consumer app with 500,000+ downloads. I fine-tune and deploy speech and language models, build retrieval-augmented generation with local LLMs, and ship the backend, data and mobile systems around them.

Experience

AI Developer · Next AI Pvt. Ltd

Nov 2025 – Present · Nepal

Own the real-time voice-to-voice AI system (STT → LLM → TTS over LiveKit and WebRTC) serving 1,000+ concurrent users, integrated into a consumer app with 500,000+ downloads.
Designed streaming orchestration in which every stage overlaps the next (partial transcripts, token streaming, first-sentence speech synthesis) to meet a < 700 ms voice-to-voice target budget.
Run it self-hosted in Nepal (FastAPI agent workers on Linux) and build its speech processing, integrations and scalable deployment workflows.
Fine-tuned Gemma 3 for customer-support replies and Piper for natural text-to-speech; deployed OmniVoice and Qwen3-TTS inference.
Built a customer-support AI agent with retrieval-augmented generation over company documents and local LLM inference, so answers are grounded in retrieved sources.
Prompt-engineered 48 English-learning AI characters for conversational practice; exploring local Nepali STT, LLM and TTS (IndicConformer, Nemotron).

Full-Stack Developer (contract) · Simal

Jun 2025 – Present · Nepal

Built an end-to-end multi-tenant SaaS mobile app with React Native, Expo, Supabase and PostgreSQL, isolating each company's data with row-level security.
Manufacturing ERP: offline-first production capture (WatermelonDB) with paged push / pull sync, and BOM-driven dynamic forms rendered from JSON blueprints.
Built configurable reports and a generative UI dashboard whose specs are validated against a component catalog before rendering.
Multi-company CRM: org-scoped TanStack Query cache, typed feature APIs with Zod, workspace roles and row-level security.
Freight dispatch (in development): load requests, truck and driver assignment, typed RPC commands with revision checks and retry IDs, and four delivery checkpoints.

Skills

Machine learning & LLMs: Large language models, RAG, embeddings, vector search, fine-tuning (Gemma 3, Piper), prompt engineering, agentic workflows, structured output, document chunking
Speech AI: Speech recognition / ASR (Whisper, Nemotron, IndicConformer), text-to-speech (Piper, Qwen3-TTS, OmniVoice, Kokoro), streaming audio, voice agents
Backend & real-time: Python, FastAPI, Flask, REST APIs, WebRTC, WebSockets, LiveKit
Infrastructure: Docker, Linux, AWS, CI/CD, Jenkins, Git, Grafana, self-hosted deployment
Data: PostgreSQL, Supabase, row-level security, MySQL, MongoDB, WatermelonDB
Product engineering: React Native, Expo, TypeScript, JavaScript, TanStack Query, Zod
Languages: Python, TypeScript, JavaScript, Java, C++, Dart

Projects

Real-time voice AI — Streaming STT → LLM → TTS pipeline on LiveKit; 1,000+ concurrent sessions, self-hosted.
Customer-support AI (RAG) — Retrieval over support documents with a fine-tuned, locally served Gemma 3.
Plant Monitoring System — ESP32 sensors with automatic watering.
PyQuest — Learning game with Gemini-generated questions.
CPR Feedback — Arduino device that graphs chest-compression depth and rate.

Certifications

AWS Certified Cloud Practitioner
AWS Machine Learning Fundamentals

Awards

NASA Space Apps Challenge 2023 — Honourable Mention
Hackathon winner

Education

BSc (Hons) Computing with Artificial Intelligence
2023 – 2026