Suman Paudel

Senior AI Engineer · Nepal

Suman Paudel — Senior AI Engineer in Nepal, Voice AI and RAG

SUMAN

PAUDEL

I design and ship production voice AI, RAG, LLM, and machine learning systems — from sub-second agents handling 2,000+ daily calls to GPU-backed retrieval platforms with 95% accuracy.

Get in touchSee the work
Scroll▼
<900ms
Latency
99.9%
Uptime
2,000+
Calls / day
95%
Accuracy
About

Five years in the trenches

Voice AI · RAG · infrastructure · production systems at scale

I'm a Nepal-based AI and machine learning engineer who builds systems that listen, understand, and respond — at scale. Currently a Senior AI Engineer at Leapfrog Technology, where I architect voice agents handling 2000+ daily calls with sub-second latency, and build RAG systems that reason across millions of manufacturing documents.

Five years across data science, machine learning, and voice engineering. From fine-tuning vision models to orchestrating multi-region LiveKit infrastructure. From winning international NLP competitions to shipping production healthcare AI.

Languages:
PythonSQLRTypeScriptGoRust
ML / DL:
PyTorchTensorFlowscikit-learnHuggingFacepandasnumpyJAX
LLMs & AI:
LangChainLlamaIndexCrewAILangGraphGPT-OSS-20BQwen3LlamaMixtralClaudeGemini
Voice AI:
LiveKitDeepgramWhisperElevenLabsCartesiaWebRTCSIP TrunkingTwilioTelnyxPipecat
RAG & Search:
FaissChromaPineconeMilvusElasticsearchBM25Hybrid SearchRe-ranking
Infra & MLOps:
AWS (EC2/EKS/ECS)AzureAlibaba CloudKubernetesHelmDockervLLMSGLangGitHub ActionsMLflow
Databases:
PostgreSQLMongoDBInfluxDBRedisSupabaseSQLite
Senior AI Engineer
Leapfrog Technology, Inc
Apr 2025 — Present
  • Manufacturing RAG + SQL Agent: vLLM platform with GPT-OSS-20B, Qwen3-4B embeddings, Qwen3-VL-8B on H100/A100/A40 GPUs. LangGraph orchestration. 95% retrieval accuracy, 78% complex reasoning.
  • Document Processing Engine: Fine-tuned DeepSeek OCR, PaddleOCR, OLMO-v2, Qwen3-8B OCR using SFT/DPO. Improved accuracy from 84% to 95% on 1000+ labeled samples via Gemini 3.0 teacher.
  • Real-Time Voice Agent: Self-hosted LiveKit on AWS, 2000+ daily calls, <900ms latency, 99.9% uptime. Multi-channel SIP across Twilio, Plivo, Telecimi, Acefone with auto-failover. 60% failed call reduction.
  • Qwen3 TTS: Open-source TTS via Alibaba Cloud, 70% cost reduction. Healthcare RAG with 92% query accuracy, 40% medication adherence improvement.
AI Engineer
Dogma Group
Apr 2024 — Apr 2025
  • RAG-Based Tender Processing System: LangChain + OpenAI/Gemini + BM25/MMR reranking + MongoDB memory, deployed on Docker with Azure. 30% reduction in manual effort.
Data Scientist
BitsKraft
Apr 2021 — May 2024
  • Credit Risk Modelling: Improved accuracy from 54% to 71%. Azure + FastAPI backend for real-time scoring.
  • Customer segmentation & churn prediction pipelines for enterprise clients.
Master in Data Science
Tribhuvan University
2024 — 2026 · GPA 3.7/4.0
B.Tech in Computer Science
Chandigarh University
2017 — 2021
01LATENCY IS EVERYTHING

Every millisecond is a conversation lost

I architected self-hosted LiveKit infrastructure on AWS achieving
<900ms end-to-end latency, orchestrating 2000+ daily calls at
99.9% uptime. Because humans don't wait for machines.
p50_latency_ms=871
concurrent_calls=120
02OWN THE STACK

Open source is not a compromise. It's a weapon.

Integrated open-source Qwen3 TTS reducing costs by 70%.
Fine-tuned OCR/VLM models (84% → 95%) on custom datasets.
Self-hosted vLLM on H100 serving GPT-OSS-20B at production scale.
tts_cost_delta=-70%
gpu=H100
accuracy=95%
03RETRIEVAL IS REASONING

Ground every answer in truth

Designed RAG + SQL agent pipelines solving large-scale manufacturing
analytics with LangGraph. 95% retrieval accuracy. 78% complex reasoning.
Won CHIPSAL @ COLING 2025. ACL-published research.
retrieval_acc=95%
multi_step=78%
competition=1st place
AI / ML + RAG

Retrieval and reasoning

Production RAG systems · LLM fine-tuning · NLP research · vLLM inference

Manufacture-Excel-RAG
Production RAG system for manufacturing Excel analytics. Multi-agent SQL query generation, hybrid search, GPU-accelerated vLLM inference on H100/A100. Docker, Alembic, comprehensive evaluation framework.
PythonFastAPILangGraphvLLMDockerPostgreSQL
SQL-GRINDA
Elasticsearch-based SQL AI Agent for Korean manufacturing enterprises. Full RAG pipeline with FastAPI, Alembic migrations, evaluation datasets, and React frontend.
PythonFastAPIElasticsearchDockerAlembic
uvc-text2sql
Text-to-SQL agent for manufacturing data across PostgreSQL + InfluxDB. LangGraph orchestration with LangSmith tracing. Vercel proxy, Docker deployment, comprehensive eval datasets.
PythonFastAPILangGraphPostgreSQLInfluxDBDocker
counselor-data-pipeline
RAG pipeline for AI Counselor system. Document processing, embedding, hybrid retrieval with evaluation framework. Built cost model analysis for RAG infrastructure.
PythonFastAPIRAGDockerAlembic
hwpx-intelligence
Korean HWPX document parsing & RAG pipeline. Custom document structure extraction, table parsing, and intelligent chunking for Korean regulatory documents.
PythonRAGDocument AIOCR
pps-mono-repo
AI-powered proposal generation & evaluation system. Multi-service architecture with AI engine, analysis pipeline, knowledge hub, and voice evaluation via Pipecat.
PythonFastAPIDockerPipecatRFP
Voice Agents

Real-time conversation

LiveKit · SIP trunking · Real-time speech · 2000+ calls/day · 99.9% uptime

eugenix-voice-agent
Hair transplant clinic voice AI handling 2000+ daily calls. Multi-persona agent system (Welcome/Closer), LiveKit transport, SIP trunking across 4+ carriers, 99.9% uptime.
PythonLiveKitSIPDockerAlembic
voice-agent
Core voice AI agent platform with call dispatcher, noise filtering, STT/TTS/LLM factory. PM2 process management, egress template generation. Production-hardened.
PythonLiveKitDeepgramPM2Docker
callpal-livekit-py
AI phone calling platform with LiveKit backend. SIP trunk integration, Flask-like app structure, docker-compose for one-command deploy.
PythonLiveKitSIPDocker
chattrvox-production-saas
Production voice chat SaaS with React/Vite frontend, Node backend, Supabase auth, and LiveKit agents. Full production/staging Docker environments.
ReactNodeSupabaseLiveKitDocker
AIRAZOR Voice Agents
Multi-agent voice platform with control plane, frontend, and multiple Python agent backends. Monorepo with shared runtime core. Docker swarm deployment.
PythonLiveKitReactDockerMakefile
docuvoice
Document-to-voice platform with backend, frontend, agents, and infra services. Docker-compose for full-stack deployment.
PythonTypeScriptDockerVoice AI
rapida-voice-ai / voice-ai
Go-based voice AI platform with LiveKit integration, Pipecat end-of-speech detection (ONNX), document RAG, and Go LLM agent executor.
GoLiveKitPipecatONNXDocker
kai-voice / moshi-livekit / zenvoice
Additional voice platforms: Kai Voice (multi-agent + token server), Moshi + LiveKit bridge, ZenVoice (NestJS + Next.js + Python agent stack).
PythonLiveKitNestJSNext.jsDocker
Infrastructure

Systems that stay up

Kubernetes · LiveKit self-hosting · AWS/Azure · Multi-region · Helm · Terraform

k8s-deploy
Kubernetes deployment infrastructure with Helm charts and Terraform. Production LiveKit K8s values including Redis, Traefik ingress, and autoscaling configuration.
KubernetesHelmTerraformAWS
livekit-infra-deployment
Self-hosted LiveKit infrastructure on AWS/Azure. Configs for multiple regions (AMS, Miami), Docker-compose + Kubernetes manifests, cloudflared tunnels.
LiveKitAWSAzureDockerKubernetes
oneinbox-lk-infra
Multi-region LiveKit deployment for OneInbox platform. SIP trunking across 4 carriers, auto-failover, cloudflared tunneling, production monitoring.
LiveKitSIPAWSMonitoring
AWS Setup Scripts
Automated LiveKit deployment scripts for AWS (EC2/EKS). Eugenix and Aivoco production environment setup with security hardening and monitoring.
AWSBashLiveKitTerraform

Let's build
what matters.

The future belongs to those who ship outcomes, not outputs.
If you're building something that listens — I want to help.

EmailLinkedInGitHub

Privacy · Terms