Bengaluru, India
I build AI systems
end to end.
Software engineer at A.P. Moller Maersk. I work across data pipelines, backend systems, frontends, and AI and I care as much about the business problem as the technical solution. So far that has meant turning days-long manual processes into tools people actually rely on.
01 · ABOUT
A bit about me.
I graduated from Symbiosis Institute of Technology in 2025, joined Maersk as an intern, and converted to full-time after a year. Since then I've worked across data engineering, backend, frontend, and AI — often all on the same project.
I've also published two research papers in Scopus-indexed journals — on multimodal RAG systems and computer vision for agriculture. Research has a different kind of rigour than production engineering, and I find it useful to keep both.
Career
02 · WORK
Work Experience
Three tools live in production with real users — full ownership from database schema to CI/CD.
AI Analytics Copilot
Agentic analytics copilot for an enterprise logistics operations team. Orchestrates 4 LLM workflows to answer natural-language questions about operational cost data over a cloud data warehouse — cutting analysis from hours to seconds.
- FastAPI
- LangGraph
- LLM
- React
- PostgreSQL
- Docker
- AKS
InvoiceFlow
Exception review tool for a regional finance operations team. Surfaces ~750 invoice exceptions daily from a high-throughput data pipeline across regional offices — letting ops teams review, action, and sync decisions back upstream automatically.
- FastAPI
- Vue 3
- PostgreSQL
- Databricks
- Azure AD
- Docker
Allocation Engine as a Service
Finance allocation rule management platform for an enterprise finance ops team. Replaces manual spreadsheet workflows with a structured web interface for creating, browsing, updating, and simulating allocation rules across P&L categories.
- FastAPI
- Vue 3
- Pinia
- SQL Server
- Azure AD
- Vitest
- Cypress
03 · PROJECTS
Projects
ML research, full-stack apps, mobile, and cybersecurity — two with Scopus-indexed publications.
PromptEase
Evaluated 5 RAG architectures across 17 pipeline configurations for multilingual product support, combining BM25, dense retrieval (DPR, FAISS), and a ColPali multimodal approach. Achieved 94% Top-1 retrieval across 170 engineering manuals. Published in IAJIT (2026).
- FastAPI
- LangChain
- ColPali
- FAISS
- IndicBERT
- mT5
- Gemini
Dual-module ML system for PE malware and polymorphic URL detection using ensemble classifiers (Random Forest, Gradient Boosting) with TF-IDF and SMOTE resampling. Evaluated on 420K+ URLs. Published in IJCDS (Scopus-indexed, CiteScore 2.3).
- Python
- scikit-learn
- Random Forest
- Gradient Boosting
- TF-IDF
- SMOTE
CNN transfer learning platform for crop risk assessment using smartphone images. Benchmarked 7 architectures — EfficientNet-B0 achieved 92%+ accuracy on crop species and growth stage detection from a 14K+ image dataset.
- Python
- fastai
- EfficientNet-B0
- VGG16
- Transfer Learning
- React
- Flask
Farm Policy Advisor
Multilingual AI chatbot helping Indian farmers discover government schemes in Hindi and regional languages, with voice interaction, eligibility logic, and form auto-fill. Selected for the FAER Integra Scholar Program 2023-24.
- Android
- Kotlin
- Generative AI
- NLP
- Firebase
- Python
04 · SKILLS
Tech I work with
Languages
AI & GenAI
Backend & Databases
Frontend
Data Engineering
DevOps & Cloud
05 · RESEARCH
Publications
Optimising Multimodal RAG Systems for Multilingual Product Support
IAJIT (Scopus-indexed)
Evaluated 5 RAG architectures across 17 configurations for multilingual product support. A ColPali-inspired multimodal approach achieved 94% Top-1 retrieval accuracy.
Malware Detection and Classification Using Machine Learning Techniques: An Implementation Study
IJCDS (CiteScore 2.3, Q3)
Developed ML-based detection models for polymorphic URLs and PE malware, achieving 97.85% and 99.4% accuracy using TF-IDF/Logistic Regression and Random Forest with oversampling techniques.
Contact me
Please contact me directly at krishajoshi.work@gmail.com or through this form.