Agentic systems
Tool-calling agents with reflection loops, orchestration and hard limits on retries, budgets and timeouts. Built on Claude and OpenAI, mostly through AWS Bedrock.
AI Engineer · Kathmandu, Nepal
Sushank Ghimire
I build LLM agents that call tools and drive browsers, retrieval systems that survive messy documents, and the FastAPI backends that keep them at 99.9% uptime.
Now: Software Development Engineer II at Renegade InsuranceClaude · Bedrock · RAG · Python
embedding space
4,096 vectors · cos sim > 0.82
state: clustering
- userPull the deductible and effective date from this 41 page policy PDF.
- planDocument is scanned. Run OCR, locate the declarations page, then extract typed fields.
- tool_callocr_document(pages="all", engine="layout")
- observation41 pages, 2 tables detected, declarations page = 3
- tool_callextract_fields(page=3, schema=PolicyDeclarations)
- observation{ deductible: "$2,500", effective_date: "2026-03-01", confidence: 0.97 }
- reflectDeductible matches the premium table on page 7. No schema mismatch. Done.
- finalDeductible $2,500, effective 1 March 2026. Source: page 3, verified against page 7.
00 · What I build
Four things I do well, all of them in production
Not demos. Systems that real analysts, agents and customers depend on every day.
Retrieval and document intelligence
RAG pipelines for messy HTML and scanned documents: structure-aware chunking, hybrid retrieval, re-ranking, and multimodal extraction with OCR.
LLM browser automation
Agents that read the live DOM and complete multi-step workflows in web apps that have no API.
Production backends
FastAPI and Django services with observability, caching and error handling that hold 99.9% uptime under real traffic.
01 · Selected work
Production AI systems, with the numbers that mattered
Three systems from the last two years. Client details stay private, the engineering does not.
Case study 01Verisk Nepal
Agentic document intelligence on AWS Bedrock
Insurance analysts needed answers grounded in thousands of HTML documents, and the first chatbot was slow and often wrong.
- Claude on Bedrock with tool calling, retrieval, reflection loops and an orchestrator that routes between specialised agents.
- A preprocessing pipeline for HTML: boilerplate removal, structure-aware chunking and metadata extraction.
- Hybrid retrieval (lexical plus dense) with re-ranking, response caching and parallel tool execution.
- A FastAPI redesign with per-tool deadlines, retry budgets, schema validation and tracing.
- AWS Bedrock
- Claude
- FastAPI
- PostgreSQL
- Python
- faster and more accurate
- 2x
- increase in user retention
- 80%
- uptime after the redesign
- 99.9%
Case study 02Renegade Insurance
LLM agents that operate insurance web portals
Insurance operations run on third-party portals with no APIs. Every quote, policy lookup and document pull was a person clicking through forms.
- Browser automation agents that plan, act and observe in a loop, reading the live DOM instead of relying on brittle selectors.
- Multimodal extraction that combines OCR with vision-capable LLMs to turn scanned forms and tables into typed records.
- Agentic SEO pipelines for keyword research, competitor analysis and metadata generation.
- Python
- Playwright
- Multimodal LLMs
- OCR
- FastAPI
- workflows completed unattended
- Multi-step
- data from unstructured documents
- Structured
Case study 03Naamche (reAlpha)
Conversational AI for real estate
A US real estate platform wanted assistants that could answer buyer questions and read listing documents at scale.
- Chatbots and agentic workflows on OpenAI and Claude via Bedrock, each tuned to a client use case.
- Document extraction and parsing with NLP to automate data entry from listings and contracts.
- FastAPI and Django services that exposed the models as reliable, versioned endpoints.
- OpenAI
- Claude
- FastAPI
- Django
- PostgreSQL
- assistants across client accounts
- Scalable
- document processing
- Automated
02 · Projects
Open source and side projects
Things I built to learn, to ship, or because nobody had built them yet.
All projects
2023AI / ML
EZ-Crossword
Solves newspaper crosswords and generates new ones with transformer QA models
- Python
- PyTorch
- Transformers
- Django
- FastAPI
- Bootstrap
2023AI / ML
Automatic License Plate Detection
YOLO based detection of plates and the characters on them, wrapped in a web app
- Python
- YOLO
- OpenCV
- PyTorch
2025AI / ML
FastTrackV2
Django 5 and DRF rebuild of an engineering institute's platform with Docker, uv and a justfile
- Python 3.13
- Django 5.1
- Django REST Framework
- PostgreSQL 17
Also built
Web platforms and backends I have shipped along the way
- Info-Bank, Bagmati ProvinceDocument repository for government organisations across Bagmati Province, NepalDjango / PostgreSQL / Nginx
- Tourist-Info, Lumbini ProvinceGuide site for tourists navigating Lumbini Province, designed for large information screensDjango / PostgreSQL / JavaScript
- Fast Track Engineering InstituteThe institute's public website and backend, which I still run as Technical HeadDjango / PostgreSQL / Nginx
- Digital PianoA playable piano in C++ and SFML with single notes and common chordsC++ / SFML
03 · Research
Classical ML, done properly, with the paper to prove it
Six studies from my engineering degree. Each has a notebook you can run and a write-up you can read.
Read the studies- PCAWine, Iris
A comparative study of Principal Component Analysis on different datasets
How much variance you can throw away before class separation collapses, measured on two classic datasets.
- Decision treesWisconsin Breast Cancer
Classifying breast cancer as benign or malignant with decision trees
Gini impurity against entropy as the split criterion, with a look at depth and overfitting.
- Naive BayesObesity levels
Obesity prediction with a Naive Bayes classifier
A Gaussian and hybrid Naive Bayes approach on mixed numeric and categorical features.
- KNNStudent dropout
Student dropout prediction with K-Nearest Neighbours
Sweeping k and distance metrics to predict which students are likely to leave.
- ANNMNIST
Digit classification with artificial neural networks
A plain feed-forward network first, then the same network with regularization to see what actually helps.
- KD-treeSynthetic and tabular
Speeding up KNN with KD-trees
Replacing brute-force neighbour search with a KD-tree and measuring where it pays off.
04 · Experience
Where I have worked
From backend developer to AI engineer, with a lot of shipped software in between.
- Jul 2026 to presentcurrent
Software Development Engineer II at Renegade Insurance
Agents that operate real web applications and read real documents for insurance operations.
- Built production browser automation and RPA agents driven by LLM action loops with dynamic DOM handling, so multi-step web workflows run without a human at the keyboard.
- Developed end-to-end document parsing and extraction pipelines using multimodal LLMs and OCR to pull structured data out of unstructured documents and tables.
- Automated SEO work end to end with agentic flows for keyword research, competitor analysis, and metadata and content generation.
- Jul 2025 to Jul 2026
Software Engineer II at Verisk Nepal
Agentic systems on AWS Bedrock with Claude for client-specific insurance workflows.
- Built production agentic systems on AWS Bedrock using Anthropic Claude with tool calling, RAG, reflection loops and agent orchestration.
- Designed preprocessing pipelines for HTML documents with smart chunking and metadata extraction to feed high-quality RAG datasets.
- Doubled chatbot and agent performance on both latency and accuracy through prompt optimization, hybrid retrieval with re-ranking, caching and parallelization. User retention rose 80%.
- Found and fixed critical backend gaps (rate limits, timeout cascades, tool schema mismatches) and redesigned the FastAPI services with proper error handling and observability, reaching 99.9% uptime.
- Sep 2024 to Jul 2025
Applied AI Engineer at Naamche Inc. (reAlpha)
Conversational AI and document systems for a US real estate technology company.
- Designed conversational AI systems on OpenAI APIs and Anthropic Claude via AWS Bedrock, shipping chatbots and agentic workflows tailored to each client.
- Engineered document extraction and parsing systems that used NLP techniques to automate data processing.
- Built the FastAPI and Django backend services that put those models in front of production traffic.
- Mar 2024 to Sep 2024
AI Developer at ICEBRKR, Virtly
Research and prototyping across computer vision and NLP.
- Researched and prototyped AI solutions for business problems in natural language processing and computer vision.
- Integrated CV and NLP components into products for real-time data analysis and user interaction.
- Jan 2024 to presentcurrent
Technical Head at Fast Track Engineering Institute
Owns the institute's web platform and infrastructure.
- Designed and built the institute's website and secure Django backend from scratch.
- Cut latency by 20% through query optimization, caching and a cleaner deployment pipeline on the VPS.
- Feb 2023 to Feb 2024
Backend Developer at Perigee Solutions
Backend for payroll systems, admin dashboards and business sites.
- Built payroll management, admin dashboards and business websites with user management, reporting and workflow automation.
- Integrated REST APIs with frontend teams and improved performance, security and reliability across multiple client projects.
Education
Bachelor of Engineering in Computer Engineering
Thapathali Campus, Institute of Engineering, Tribhuvan University. Kathmandu, Nepal.
05 · Stack
Tools I reach for
The current set. It changes when something better shows up and proves itself.
AI and ML
- Anthropic Claude
- OpenAI
- AWS Bedrock
- LangChain
- LlamaIndex
- PyTorch
- TensorFlow
- scikit-learn
- YOLO
- Hugging Face Transformers
Backend
- FastAPI
- Django
- Django REST Framework
- PostgreSQL
- MongoDB
- Redis
- Alembic
Languages
- Python
- TypeScript
- JavaScript
- SQL
Infrastructure
- Docker
- Nginx
- AWS
- GitHub Actions
- Cloudflare
- Linux VPS
Practices
- Hybrid retrieval and re-ranking
- Evaluation sets
- Observability
- Prompt optimization
- Caching and parallelization
06 · Writing
Notes from building agents and retrieval systems
Longer write-ups on what broke, what fixed it, and what I would do differently.
All postsMaking tool-calling agents reliable on AWS Bedrock
Rate limits, timeout cascades and tool schema drift took down our Claude agents more than any bad prompt did. How we fixed each one and reached 99.9% uptime.
RAG on messy HTML: preprocessing, structure-aware chunking and hybrid retrieval
Fixed-size chunking fails on real documents. What moved retrieval quality for us: boilerplate stripping, heading-aware chunks, hybrid search and a re-ranker.
07 · Contact
Working on agents, retrieval or document intelligence? Let's talk.
I am Sushank Ghimire, an AI engineer based in Kathmandu and working with teams around the world. I answer email within a day, and I am happy to look at a hard problem before we talk about anything else.
Kathmandu, UTC+5:45