Skip to content
In Production

℀ Curriculum vitae

Rajan Devkota

Kathmandu, Nepal

Senior AI engineer with 3+ years building and operating production AI systems, currently leading a team of five and owning the stack end to end: document intelligence and RAG pipelines, multi-agent orchestration, vLLM serving on Kubernetes, and GPU cost engineering. Depth in high-throughput inference (vLLM, TGI, Triton) and parameter-efficient fine-tuning (QLoRA), delivered across AWS, Azure, GCP and DigitalOcean.

Download PDFEmail meLinkedIn

§ Experience

Where I have worked

Nov 2025 — Present

Senior AI Engineer

Wildenova · Texas, USA · Remote

  • Own DevOps, MLOps and AI engineering end to end while leading a team of five across four services, from Kubernetes manifests through to shipped product.
  • Architected a two-stage hybrid document pipeline that decouples layout from transcription: a deterministic layout detector (PP-DocLayoutV3 via the GLM-OCR SDK) segments and crops semantic regions — paragraphs, tables, formulas, charts — then a compact vision-language decoder (Qwen 3.5 4B) served on vLLM transcribes each high-resolution crop in context.
  • Built long-horizon document workflows on LangChain DeepAgents using planning, sub-agent delegation and a shared virtual filesystem, keeping multi-step extraction jobs coherent across 500 documents.
  • Re-engineered the serving path onto A40-class GPUs at roughly one-sixth the hourly rate of H100, holding throughput with continuous batching and Redis-backed queuing across a Rust and Python service boundary.
  • Built serverless, demand-driven inference on DigitalOcean Kubernetes (DOKS) behind an Nginx ingress and load balancer, with gRPC services, KEDA autoscaling across GPU droplets and CI/CD pipelines driving build and rollout.
  • Engineered a Graphiti-style temporal memory layer on PostgreSQL and Qdrant over end-to-end encrypted storage, cutting the in-memory index 32× with binary quantisation (float32 to 1 bit per dimension) while preserving recall through rescoring, and scaled the vector store horizontally with sharding and replication.

Jan 2025 — Feb 2026

AI Engineer

Herix Co · Tokyo, Japan · Part-time, remote

  • Designed a cost-effective due-diligence application reaching 95% accuracy across more than 500 annual reports, through optimised LLM orchestration and efficient data processing on Azure Functions.
  • Developed deep research agents for due diligence using LangGraph, improving the grounding of analytical insights.
  • Built an enterprise hybrid RAG system combining vector databases with knowledge graphs, improving precision and enabling multi-hop querying.
  • Optimised Document Intelligence model deployment with Triton Inference Server on Azure spot instances and Functions, cutting inference cost by 50%.
  • Delivered across a multi-cloud footprint, running managed model workflows on Vertex AI and event-driven document processing on Azure Functions.

Sep 2023 — Nov 2025

AI Engineer

Frost Digital Ventures · Kathmandu, Nepal · Onsite

  • Designed and built an end-to-end chatbot system using advanced retrieval and vector databases, serving over 5,500 case studies and documents to 120+ users a day.
  • Built a Nepali-language speech-to-text application by fine-tuning Wav2Vec2 on a custom audio dataset, reaching production-ready transcription accuracy.
  • Architected server infrastructure and optimised resource allocation for deploying open-source LLMs, using TEI for embeddings and vLLM for model serving.
  • Developed and managed multi-agent systems with LangGraph, integrating agentic frameworks from LlamaIndex and LangChain for complex task orchestration.
  • Fine-tuned Llama 3 (8B) with QLoRA for domain-specific adaptation, and instruction-tuned GPT-4o-mini on GPT-4o-generated data for cost-efficient inference.

Jul 2023 — Sep 2023

Data Scientist Intern

iNeuron.ai · India · Remote

  • Designed and delivered an end-to-end computer-vision project using transfer learning, deployed with FastAPI for production use.
  • Containerised and deployed applications on AWS with Docker and GitHub Actions for automated CI/CD.
  • Implemented MLOps practices including model versioning, automated testing and monitoring across the model lifecycle.

§ Education

How I trained

Jan 2018 — Jun 2023

Bachelor of Engineering, Computer Engineering

IOE Pashchimanchal Campus, Tribhuvan University · Nepal

  • Graduated First Division.

§ Skills

What I work with

AI / ML frameworks

  • PyTorch
  • Transformers
  • LangChain
  • LangGraph
  • DeepAgents
  • LlamaIndex
  • vLLM
  • TGI
  • Ollama

Serving & optimisation

  • vLLM
  • Triton
  • TensorRT
  • KEDA autoscaling
  • Continuous batching
  • QLoRA
  • Binary quantisation

Vector databases

  • Qdrant
  • Weaviate
  • Pinecone
  • FAISS
  • ChromaDB

Data processing

  • PDF / DOCX parsing
  • OCR
  • Document layout detection
  • Excel processing
  • Web scraping

Vision & NLP

  • Transfer learning
  • Vision-language models
  • Stable Diffusion
  • OCR optimisation
  • Speech recognition

Backend

  • Python
  • Rust
  • FastAPI
  • gRPC
  • PostgreSQL
  • MySQL
  • MongoDB
  • Redis
  • Celery
  • Alembic

Infrastructure & CI/CD

  • Docker
  • Kubernetes
  • Nginx
  • GitHub Actions
  • Apache Airflow
  • Terraform

Observability

  • Prometheus
  • Grafana
  • OpenTelemetry
  • Sentry
  • Arize Phoenix
  • LangSmith

Cloud

  • AWS (EC2, SageMaker, Lambda, S3)
  • Azure (VM, Blob, Functions)
  • GCP (Vertex AI)
  • DigitalOcean (DOKS, Droplets)