AI Engineer / RAG Systems / Kathmandu

RishavUpadhaya

I build production Generative AI: RAG pipelines, multi-agent workflows, and LLM backends that move from prototype to shipped systems.

View workStart a conversation
55%LLM cost reduction
96%OCR data fidelity
<4sVector search latency
1/20Leapfrog SPP cohort

01 About

Built for production

Inspired by the bold motion and visual confidence of the reference, this portfolio translates that energy into an AI engineering world: retrieval, agents, observability, and shipped product systems.

I turn fuzzy AI ideas into systems that can be measured, monitored, and shipped.

At AsterGaze, I design multi-agent workflows and RAG systems for a study-abroad CRM platform, combining LangGraph orchestration, FastAPI APIs, PostgreSQL, pgvector, OpenAI embeddings, and LangSmith observability.

I care about the engineering details that make AI useful: retrieval quality, latency budgets, token cost, schema design, evaluation, and graceful failure modes. Beyond work, I have led student communities, organized workshops, and competed in national hackathons.

26GB
Model handled on a 16 GB machine
3s
OCR batch processing from 8 hours
500+
Applicants in Leapfrog selection pool
200+
Workshop participants reached

02 Experience

Build history.

A compact timeline of the places where I have built backend systems, AI pipelines, and production-facing workflows.

May 2025 - Present

AsterGaze Technologies

Kathmandu, Nepal / Part-time

AI Engineer

  • Designed a multi-agent AI system for a study-abroad CRM platform using LangGraph and FastAPI.
  • Built hybrid RAG over PostgreSQL and pgvector HNSW, reducing vector search from about 10s to under 4s.
  • Reduced LLM inference cost by up to 55% through prompt design, context compression, and token optimization.
  • Tracked traces, latency, and failure points with LangSmith for better production visibility.

Nov 2025 - Feb 2026

Proshore.eu

Backend Developer Intern

Backend Developer Intern

  • Built a multi-engine OCR pipeline with AWS Textract, Azure Document Intelligence, and Google Vision.
  • Cut processing time from 8 hours to about 3 seconds per batch while preserving 96% data fidelity.
  • Improved OCR quality by 35% with OpenCV preprocessing for skew, grayscale, and orientation correction.
  • Developed REST APIs with Pydantic validation, error handling, and OpenAPI documentation.

Apr 2025 - Oct 2025

Leapfrog Technology

Student Partnership Program

Student Partner

  • Selected from 500+ applicants for a competitive six-month software engineering program.
  • Led backend development for Reviso.ai, an LLM-powered exam generation and evaluation platform.
  • Integrated GPT-4, LangChain, Pinecone, and FastAPI with CI/CD and collaborative code review practices.

03 Projects

Machine work.

Featured systems across low-level inference, RAG architecture, LLM products, and classical ML.

Showpiece2026

SWLP Sliding Window Layer Pipeline

Memory-efficient FP16 inference engine that runs a 26 GB transformer model on a 16 GB machine at 1.7 GB peak RAM without quantization or quality loss.

26GB → 1.7GBModel to peak RAM
2x fasterCompared with AirLLM
0.28sFlat sweep cost
Read write-upGitHub
RAG2026

JobRAG

Hybrid search over 1,000+ job listings with LangGraph state routing, pgvector HNSW, Jina reranking, and pluggable embedding providers.

1,000+Listings indexed
4-nodeAgent graph
Code
LLM app2025

Reviso.ai

Full-stack exam lifecycle platform with LLM-generated question sets, automated answer evaluation, flashcards, quizzes, and async batch assessment.

4-domainSubject RAG
4-signalProctoring
GitHub
ML2026

Diabetes Insight

Three-class diabetes prediction pipeline with supervised classification, K-Means clustering, PCA visualization, EDA, and model evaluation.

3-classPrediction
ROC-AUCEvaluation
Code

05 Stack

Tools in the stack.

The current technical toolkit I reach for when building retrieval, agentic, backend, and ML workflows.

AI and ML

LLMsRAGLangChainLangGraphLangSmithNLPFine-tuning

Backend

PythonFastAPIPydanticSQLAlchemyDockerLinux

Data

PostgreSQLpgvectorPineconePandasNumPyscikit-learn

Vision and OCR

OpenCVAWS TextractAzure Document IntelligenceGoogle Vision API

Evaluation

PrecisionRecallF1ROC-AUCCER/WERTracing

Practice

AgileCode reviewCI/CDDocumentationObservability

06 Awards

Proof under pressure.

Competitive builds and leadership work that shaped how I collaborate, prototype, and communicate when the problem is still messy.

Winner

Locus Niural AI Hack-a-Week

AIDEA Hackathon: Runner Up

Built from a whiteboard idea to a demo-ready product in 72 hours, supported by mentors from Draper Startup House Nepal and US Embassy Nepal.

Hult Prize Campus Director

Led and mentored 15+ students at Hult Prize @ Samriddhi after previously placing as 2nd runner-up as a participant.

Samriddhi IT Club

Organized 10+ workshops and community events with 200+ participants.

More hackathons

Competed in SecurityPal, SXC Sandbox, CodeYatra, AIDEA, and Locus, building prototypes in public with tight deadlines and real feedback.