Suman Paudel

Suman Paudel

Senior AI engineer in Nepal. I design and ship production voice AI, RAG and LLM systems, from sub-second phone agents handling 2,000+ calls a day to GPU-backed retrieval at 95% accuracy.

Get in touchSee the work
30+
Production systems
<900ms
Voice latency
99.9%
Uptime
2,000+
Calls / day
About

Five years in the trenches

Voice AI · RAG · infrastructure · production systems at scale

I'm a Nepal-based senior AI engineer who builds systems that listen, understand, and respond at scale. I have shipped 30+ production AI systems: voice agents handling 2,000+ calls a day at sub-second latency, retrieval and SQL agents that reason over large document and data estates, and fine-tuned document AI.

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
30+ production AI systems shipped
Apr 2025 to Present
  • Analytics agents: text-to-SQL and RAG over spreadsheets, relational and time-series data, served from self-hosted vLLM (GPT-OSS-20B, Qwen3 embeddings, Qwen3-VL) on H100/A100/A40 GPUs. 95% retrieval accuracy, 78% on multi-step reasoning.
  • Document AI: fine-tuned OCR and vision-language models with SFT and DPO on 1,000+ labeled samples. Field accuracy up from 84% to 95%.
  • Real-time voice agents: self-hosted LiveKit handling 2,000+ calls a day at under 900ms latency and 99.9% uptime. Multi-carrier SIP with automatic failover cut failed calls by 60%.
  • Open-source TTS in production at 70% lower cost. Healthcare RAG assistant at 92% query accuracy.
AI Engineer
Dogma Group
Apr 2024 to 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 to 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 to 2026 · GPA 3.7/4.0
B.Tech in Computer Science
Chandigarh University
2017 to 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 industrial
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

Industrial analytics agent
Text-to-SQL and RAG agent over spreadsheets, relational and time-series data. Multi-agent query planning, hybrid search, self-hosted vLLM on H100/A100 and a full evaluation harness. 95% retrieval accuracy.
PythonFastAPILangGraphvLLMPostgreSQLInfluxDB
Document intelligence pipeline
Fine-tuned OCR and vision-language models plus structure-aware parsing, table extraction and chunking for long regulatory documents and scanned forms. Field accuracy from 84% to 95%.
PythonPaddleOCRQwen3-VLSFT / DPORAG
Proposal automation platform
Multi-service engine that generates and evaluates long-form proposals from a knowledge hub, with an analysis pipeline and voice-based review.
PythonFastAPIPipecatDocker
Retrieval cost and quality lab
Hybrid retrieval pipeline with an evaluation framework and a per-query cost model, used to size RAG infrastructure.
PythonRAGBM25Re-rankingAlembic
Voice Agents

Real-time conversation

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

High-volume phone agent
Multi-persona phone agent (greeter and closer) handling 2,000+ calls a day. LiveKit transport, SIP trunking across four carriers with failover, 99.9% uptime.
PythonLiveKitSIPDocker
Voice agent runtime
Shared runtime behind several deployments: call dispatcher, noise filtering, pluggable STT, TTS and LLM providers, process supervision and call recording.
PythonLiveKitDeepgramPM2
Voice agent SaaS
Voice agent product with a React frontend, Node backend, Supabase auth and LiveKit agents, run as separate staging and production stacks.
ReactNodeSupabaseLiveKit
Go voice pipeline
Go agent executor with LiveKit transport, ONNX end-of-turn detection and document RAG.
GoLiveKitPipecatONNX
Infrastructure

Systems that stay up

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

LiveKit on Kubernetes
Helm and Terraform for production LiveKit: Redis, Traefik ingress and autoscaling.
KubernetesHelmTerraformAWS
Multi-region media infrastructure
Self-hosted LiveKit across regions on AWS and Azure with SIP failover, secure tunnels and production monitoring.
LiveKitSIPAWSAzure
GPU inference serving
vLLM and SGLang serving LLM, embedding and vision-language workloads on H100, A100 and A40 GPUs.
vLLMSGLangDockerNVIDIA
Hardened cloud setup
Repeatable AWS provisioning (EC2, EKS) for production voice deployments, with security hardening and monitoring.
AWSTerraformBashMonitoring

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