Utsav Paudel is a Machine Learning Engineer and AI Engineer based in Kathmandu, Nepal. He has 2.5+ years of experience building Generative AI systems, RAG pipelines, LLM applications and Voice AI agents, mostly on AWS Bedrock. He works at Leapfrog Technology.

Who is a Machine Learning Engineer or AI Engineer in Nepal?

Utsav Paudel is a Machine Learning Engineer and AI Engineer in Kathmandu, Nepal. He builds production LLM systems for fintech and healthcare clients: RAG chatbots, text-to-SQL assistants, GraphRAG systems and voice agents. He is open to consulting and collaboration.

What has Utsav Paudel built?

A visa analytics RAG chatbot on AWS Bedrock and Claude Haiku 4.5. GoLend AI, a text-to-SQL assistant on Amazon Nova Pro and pgvector. A voice agent for patient calls on Amazon Nova Sonic and Twilio that cut scheduling time by 60%. Lender IQ, a GraphRAG system over 10,000+ financial documents. He also maintains two open-source Claude Code tools, cc-explain-this and jev-router.

What is Utsav Paudel's tech stack?

Python, PyTorch, LangChain, LangGraph, FastAPI, AWS Bedrock, AWS Lambda, Claude, Amazon Nova, GPT, Gemini, PostgreSQL with pgvector, and Neo4j.

Where did Utsav Paudel study and work?

He has a Bachelor in Computer Engineering from Kantipur Engineering College, Kathmandu (2017–2022). He worked at E.K. Solutions and Fusemachines before he joined Leapfrog Technology in May 2025.

How can I contact or hire Utsav Paudel?

Use the contact section on this page, or send a message on LinkedIn (linkedin.com/in/utshabpaudell). He is open to consulting and collaboration on LLM, RAG and voice AI projects.

ML Engineer @ Leapfrog Technology · Kathmandu, Nepal

Utsav Paudel Machine Learning & AI Engineer

Building RAG pipelines, voice agents, text-to-SQL engines and GraphRAG systems.

I've spent the last 2.5 years putting LLM systems into production, mostly on AWS Bedrock. Lately that means chatbots that cite their sources and voice agents that pick up the phone.

See the work Download CV

Years in production AI: 2.5+
Documents in RAG & GraphRAG: 10k+
LLM extraction accuracy: 95%
Less scheduling overhead: 60%

01Selected work

Some things I've shipped at work.

Each panel replays what happens inside the system, step by step. The steps follow the real pipelines, but the questions and data are made up.

01Visas · Global staffing2025 · Leapfrog

Visa analytics chatbot

A RAG chatbot on AWS Bedrock and Claude Haiku 4.5 that answers visa questions for a global staffing team. The model never writes a link itself. It cites short IDs like [1], and once the answer is done, an ID-mapping step swaps each ID for the source URL stored in that chunk's metadata. So it can't make one up.

Role: Lead ML Engineer
Status: In production

AWS Bedrock
Claude Haiku 4.5
RAG
Python
FastAPI

visa-analytics / rag.rundone · 3 sources linked

QWhat does an H-1B visa require?

retrieve
re-rank
generate
cite-map

1uscis_h1b_overview · §20.93
2dol_lca_guide · §10.88
3h1b_cap_season · §30.81
l1_visa_overview · §20.34
o1_visa_faq · §40.21

An H-1B needs a job offer in a specialty occupation¹ and a certified Labor Condition Application from the employer². Most new petitions also go through the annual cap lottery³.

1uscis.gov/…/h-1b-specialty-occupations
2dol.gov/…/lca
3uscis.gov/…/h-1b-cap-season

02Fintech · Lending2025 · Leapfrog

GoLend AI assistant

Loan officers ask questions in plain English and get the answer in plain English. Amazon Nova Pro writes a read-only SQL query, runs it, and explains the result. To find the right tables it mixes BM25 with pgvector search (IVFFlat and HNSW indexes) and MMR re-ranking, with AWS Guardrails on top.

Role: ML Engineer
Impact: 50% faster queries

Nova Pro
pgvector
BM25
HNSW
PostgreSQL

golend / nl→sqlanswered · 4 rows · 38 ms

QWhich applicants missed their third installment?

applicants
loans
installments
branches
payments
officers

SELECT a.id, l.amount, i.due_at
FROM applicants a
JOIN loans l ON l.applicant_id = a.id
JOIN installments i ON i.loan_id = l.id
WHERE i.seq = 3 AND i.paid_at IS NULL
ORDER BY i.due_at DESC
LIMIT 50;

Four applicants missed their third installment. The latest is APL-2291, with 420,000 due on 14 Aug 2025. The other three are APL-1873, APL-3310 and APL-0954.

03Healthcare · Voice2025 · Leapfrog

Voice agent for patient calls

When a patient calls to move a follow-up, this agent answers, checks their insurance and books a new time. It runs on Amazon Nova Sonic and Twilio. Behind it, referral documents go through Textract and Nova Pro and end up in OpenEMR. Staff now spend 60% less time on scheduling.

Role: ML Engineer
Impact: 60% less overhead

Nova Sonic
Twilio
Textract
OpenEMR
AWS Lambda

patient-line / callended · 00:31 · booked

agentThanks for calling. How can I help today?
callerHi, I need to move my follow-up. My daughter is sick.
agentSorry to hear that. Dr. Mehta has Thursday at 10 or Friday at 2. Which works?
callerFriday at 2, please.
agentBooked. I'll text the confirmation to the number on file.

⏺verify_patient(caller_id)

⎿ P-4471 · insurance active

⏺find_slots(dr_mehta, 3d)

⎿ Thu 10:00 · Fri 14:00

⏺book(openemr, Fri 14:00)

⎿ appt A-2093 confirmed

⏺calendar.add(dr_mehta)

⎿ added Fri 14:00 · freed Wed 11:00

⏺send_sms(•••0193)

⎿ confirmation delivered

dr. mehtathis week

WedThuFri

9111315 P-4471 free freeP-4471

04Fintech · Analytics2024 · Fusemachines

Lender IQ (GraphRAG)

Plain vector search struggles with questions like "which guarantors co-sign loans in three or more portfolios?" because the answer is spread over many files. We used LLMs to pull entities and relationships out of 10,000+ financial documents into Neo4j, then answered by walking the graph. Accuracy went up 40%.

Role: ML Engineer
Impact: 40% accuracy lift

Neo4j
GraphRAG
LangChain
FastAPI
GPT

lender-iq / graphrag.querydone · 1 match · 6 edges

QWhich guarantors co-sign loans across 3+ portfolios?

MATCH (g:Guarantor)-[:CO_SIGNS]→(:Loan)-[:IN]→(p:Portfolio) WITH g, count(DISTINCT p) AS n WHERE n ≥ 3

A applicant · G guarantor · L loan · P portfolio

G-17 is the only guarantor who co-signs loans in three or more portfolios. Its loans L-204, L-118 and L-152 are in P-retail, P-sme and P-agri.

+Side projects

Two open-source tools for Claude Code.

claude code · ~/auth-service5 steps · 1 failed · 1 risky

> Make sessions last 30 minutes.

⏺Bash(grep -rl "SESSION_TTL" src)

⎿ src/auth/session.ts

⏺Update(src/auth/session.ts)

- TTL = 15 * 60+ TTL = 30 * 60

⏺Bash(npm test)

⎿ 2 failed, 41 passed

⏺Update(test/session.test.ts)

⎿ updated 2 lines

⏺Bash(git push origin main)

⎿ main → main

explain thisplugin

✓Bash
Searches the code for every file that sets the session timeout.
✓Edit
Changes the session timeout from 15 to 30 minutes.
✗Bashfailed
Two tests fail because the test clock still expects 15 minutes.
✓Edit
Updates the two tests to expect the new 30-minute timeout.
✓Bashrisky
Sends your commits to the shared main branch, where others will see them.

cc-explain-this

Claude Code moves fast, and some of its shell commands are hard to read at a glance. This plugin adds a side pane that explains each step in one plain sentence. It flags risky steps (deletes, pushes, installs) and says why a step failed. The explanation is written after the step is done, so Claude doesn't slow down.

TypeScript
Claude Code plugin
Claude Haiku
MIT

jev-router

A small Rust proxy that picks the model and effort level for each message you send to Claude Code. Quick questions go to Haiku, everyday work to Sonnet, and hard or risky work to Opus. It also watches the prompt cache, so it won't switch models mid-conversation when that would cost more than it saves.

Rust
LLM routing
Prompt caching
Claude Code

02Training log

Where I've worked so far.

Plotted as a loss curve, because I couldn't help it. Every new job bumps it up a little before it settles.

train/lossrun: career · still training

epoch 04May 2025 – present

Machine Learning Engineer

Leapfrog Technology Pvt. Ltd. · Kathmandu, Nepal
- Visa Analytics Chatbot on AWS Bedrock + Claude Haiku 4.5: 4-stage RAG with ID-mapping citations.
- GoLend AI: text-to-SQL on Amazon Nova Pro + PostgreSQL pgvector.
- Voice AI agent on Nova Sonic + Twilio: 60% less scheduling overhead.
- Automated patient-referral pipeline: Textract → Nova Pro → OpenEMR.
- Bedrock
- Nova Pro
- Nova Sonic
- Twilio
- Lambda
epoch 03Sep 2024 – May 2025

Associate ML Engineer

Fusemachines Pvt. Ltd.
- RAG & GraphRAG systems with a 40% accuracy improvement across 10,000+ financial documents.
- LLM extraction pipelines (GPT, Gemini) reaching 95% accuracy.
- Retrieval optimization that cut query time by 50%.
- Neo4j
- LangChain
- GPT
- Gemini
- Python
epoch 02Jan 2024 – Jun 2024

Associate ML Engineer

E.K. Solutions Pvt. Ltd.
- 5+ scalable ML APIs on FastAPI + PyTorch serving 1,000+ daily requests.
- 25% LLM accuracy lift with Chain-of-Thought prompt engineering.
- 35% faster transformer inference with 20% less memory.
- FastAPI
- PyTorch
- Docker
- Redis
epoch 01Oct 2023 – Jan 2024

ML Engineer Trainee

E.K. Solutions Pvt. Ltd.
- Trained ML models and built the pipelines around them.
- Built NLP and computer-vision prototypes.
- PyTorch
- OpenCV
- NumPy

03Stack

The tools I actually use.

ai / ml07

PyTorch
LangChain
LangGraph
LangSmith
CrewAI
Google ADK
OpenCV

models05

Claude
Amazon Nova
Nova Sonic
GPT
Gemini

languages03

Python
SQL
C++

serving03

FastAPI
Django
Django REST

data04

PostgreSQL · pgvector
Neo4j
MongoDB
SQLite

cloud06

AWS Bedrock
Lambda
EC2
SQS
SNS
Vertex AI

infra05

Docker
MLFlow
Redis
Celery
Alembic

04About

Hi, I'm Utsav.

Portrait of Utsav Paudel, Machine Learning Engineer from Nepal, in Langtang with a snowy peak behind — Langtang, Nepal

I'm a machine learning engineer and AI engineer in Kathmandu, Nepal. Most of my work is getting LLM systems to behave once real people start using them.

I like projects where one kind of input turns into another: a phone call that ends up as a database row, or a pile of PDFs you can ask questions.

These days I'm at Leapfrog Technology. I work mostly with AWS Bedrock, Claude and Amazon Nova. Before that I was at Fusemachines and E.K. Solutions, where I learned retrieval and prompt engineering, and spent a lot of time making transformer models cheaper to run.

Outside work I'm usually on a trail. Langtang and Kori are in the photos below, and I take far too many mountain pictures. If you have a messy problem with no obvious answer yet, I'd like to hear about it.

Based: Koteshwor, Kathmandu
Now: Leapfrog Technology
Focus: GenAI · RAG · Voice
Open to: Consults & collabs