Experience
Data Scientist / AI Engineer · Climate Clean Solutions & LowPropTax
- Own production for a multi-agent LLM platform that automates property-tax appeals end to end, running on AWS across development and production.
- Built the infrastructure and release practice: all infrastructure as code with Terraform, CI/CD with GitHub Actions, releases that ship exactly what was tested, and least-privilege access throughout.
- Designed the evaluation layer: a QA agent checks every valuation before a human sees it, and production outcomes are traced end to end to verify they are correct.
- Lead LLM cost and model-selection decisions across the platform.
AI/ML Research Fellow · Fusemachines Nepal
- Built an automated model-evaluation pipeline (MAE, RMSE, MAPE, SMAPE) deployed across forecasting models after review; fine-tuned and evaluated vision-language models (CLIP, BLIP, PaliGemma 2).
RPA Developer · Quickfox Consulting
- Built an LLM-powered data-cleaning pipeline for legacy medical databases; built RPA bots for data entry and web scraping.
AI Research Intern · NAAMII
- Clinical and radiomics analysis for a radiation-oncology study under Dr. Taman Upadhaya: permutation tests, ROC AUC, and ML models for biomarker signal discovery.
Projects
Zeroth: evaluation-gated clinical multi-agent platform
- Genomics, imaging, radiomics and clinical-note agents each produce findings. An evaluation gate returns PASS, FLAG or BLOCK and refuses to display any claim without provenance; synthetic data cannot reach a report unlabeled.
- Security by design: timing-safe authentication, revocable sessions, per-source throttling, a PHI allowlist so patient identifiers never enter claims, logs or prompts, and an append-only audit log.
RAG pipeline for job data retrieval · Code
- Hybrid retrieval combining BM25 sparse search with dense embeddings and Cohere reranking, served via FastAPI and containerized.
Travya: multi-agent travel companion · Code
- Research, planning and booking agents on a microservices design with streaming responses and Google Maps, Amadeus and Stripe integrations. Built during the Fusemachines fellowship.
Text summarization with LSA and T5 · Code
- Extractive and abstractive summarizer: T5 fine-tuned on PubMed alongside LSA, evaluated with ROUGE. Final-year project.