I build, evaluate and ship machine learning and LLM systems on AWS: RAG and Amazon Bedrock assistants with grounding and safety checks, privacy-preserving AI workflows for banking, time-series forecasting backtested against baselines, edge computer vision, and MLOps pipelines with gated releases, automatic rollback and drift monitoring.
Most of my open-source work tackles real problems in Nepal (air, rain, roads, power, farming and earthquakes), built on real data and measured against honest baselines.
- AI Software Engineering Intern at Asterdio, Lalitpur: AI tools for the Breqo app, 3 agentic-AI prototypes for Nepali commercial banks (KYC update, AML alert triage, credit review), and a bank-side Privacy Shield that tokenises 8 customer identifiers and blocks any upload that fails a PII leak test
- Machine Learning Intern at FlyRank AI, Kathmandu: applied ML, AI agents and AI search (SEO, GEO)
- Independent Contractor, AI Training at Mercor: rubric-based AI training and evaluation data for a leading AI research lab; member of Mercor's Machine Learning Engineer Talent Network
- 🎓 BSc (Hons) Computer Science with AI, Birmingham City University (Sunway College Kathmandu), expected 2028
Previously: Machine Learning Engineer Apprentice at TekBay Digital Solutions (full ML lifecycle on AWS, fraud-detection capstone with S3, Glue DataBrew and SageMaker) · Solutions Architect Apprentice at Adex International (Terraform infrastructure as code that cut manual provisioning time by 40%) · Data Engineer at DataCamp
| Project | What it does | Results |
|---|---|---|
| ⚡ Bijuli: Grid demand forecasting + MLOps | Day-ahead hourly electricity demand for Nepal's grid, with a model registry, shadow rollout, automatic rollback and drift monitoring | 3.27% MAPE over 23,184 walk-forward forecasts, 30% below the best naive baseline; gated releases cut worst-week error from 9.8% to 7.45% |
| 🚦 Chowk: Edge AI traffic classification | Vehicle-mix counts for Kathmandu junctions from 30 real street videos, running on the edge with a serverless AWS backend | Accuracy 63.2% → 93.9%; traffic-volume (PCU) error 45.7% → 1.3%; 24 ms per 720p frame on one CPU core |
| 🌾 Parkha: Bilingual crop pest advisor | Nepali/English Amazon Bedrock assistant that computes safe-harvest dates from official pesticide labels, never from model output | Safety gate caught 290/290 seeded wrong answers; 18/20 on a held-out evaluation set |
| 🔐 Gopya: Secure company RAG chatbot | Role-based access control enforced inside the vector query, plus PII redaction and prompt-injection guards | Right document in the top 4 for 100% of 69 test questions; 0 leaks across 45 attacks |
| 🌧️ Sodhnus: Bilingual rainfall assistant | Bedrock tool-use assistant over 5 read-only tools that answers only from verified rainfall data | A deterministic number checker caught 97% of wrong answers in a 350-case adversarial test; Guardrails cost cut 4.2× |
| 🏞️ Barsha: Monsoon rainfall lakehouse | Serverless lakehouse on S3, Lambda, DuckDB and Athena over 50 years of Nepal station data | 67,215 station-days for about $1.20/month; quality checks caught 4,599 false zero readings |
| 🌫️ Hawa: Kathmandu PM2.5 forecasting | Hourly next-day air-quality forecasts with English/Nepali alerts on serverless AWS | 15% lower error than the best naive forecast; 40% fewer missed bad-air days |
| 🆘 Surakshit: Earthquake check-in service | Multi-region, queue-first "I'm safe" check-ins by web or SMS | 0% lost in a 15-million-user load simulation, vs 45% for a naive design |
More projects
- 🤖 Bedrock AgentCore Support Agent: customer support agent on Amazon Bedrock AgentCore with Strands Agents, Cognito JWT auth, Cedar policies and OpenTelemetry tracing
- ☸️ AI Customer Service Agent on EKS: Strands Agents with self-hosted Qwen2.5-3B (vLLM + LiteLLM) on Amazon EKS
- 🧠 MLOps Depression-Risk Pipeline: Airflow, FastAPI, Redis, MariaDB, Evidently AI, Prometheus and Grafana on Docker Compose
- 🏔️ Trekker Safety Alerts: trek check-ins and storm holds for Himalayan passes using NOAA GEFS forecasts
- ♻️ Waste Flow Pipeline: lakehouse on AWS Open Data modelling Kathmandu Valley waste flows and landfill blockades
- 📂 All repositories
Machine learning: scikit-learn · LightGBM · time-series and quantile forecasting · walk-forward backtesting · computer vision (YOLOX, edge inference) · Amazon SageMaker · drift detection (Evidently AI)
Generative AI and LLMs: Amazon Bedrock (tool use, Guardrails, AgentCore) · RAG (LangChain, FAISS, Chroma, Milvus, BM25 hybrid search) · AI agents (Strands, MCP, multi-agent A2A) · LLM-as-a-Judge evaluation · vLLM · LiteLLM · Langfuse · PII masking
AWS: IAM · VPC · EC2 · S3 · Lambda · Step Functions · EventBridge · SQS · API Gateway · DynamoDB · RDS · Athena · Glue DataBrew · IoT Core · CloudFront · WAF · CloudWatch
MLOps and DevOps: Terraform · AWS SAM · Docker · Kubernetes (EKS) · GitHub Actions · MLflow · Apache Airflow · FastAPI · Prometheus · Grafana · Linux
Languages and data: Python · SQL · Bash · Pandas · NumPy · DuckDB · Parquet
AWS architecture · MLOps · forecasting · RAG · building and evaluating agents on Amazon Bedrock · making LLM answers safe to trust
Open to AI/ML engineering roles and collaborations. The fastest way to reach me is email or LinkedIn.