Spandan Mishra

Head of AI & Data Science, Palo Alto Networks

I work on continual learning and self-improving agents, and build language models for low-resource South Asian languages. My PhD work focused on applied statistics, Bayesian inference, and structural health monitoring. In 2026, I trained Ama Dablam, a 322M-parameter model for Nepali, Maithili, and Bhojpuri.

Email / Google Scholar / ORCID / GitHub / CV

Research Interests

Continual learning • Self-improving agents • Language models for low-resource languages (Nepali, Maithili, Bhojpuri) • Multi-agent LLM systems • Context engineering • Bayesian inference • Structural health monitoring

News

2026 Released Ama Dablam, a 322M-parameter language model for Nepali, Maithili, and Bhojpuri

Selected Work

Ama Dablam: A Language Model for Nepali, Maithili, and Bhojpuri

A 322M-parameter language model trained on 85% Devanagari, 5% IAST, and 10% phonetic romanization. Achieves 0.4061 bits-per-byte on held-out Nepali text.

Technical Report

agent-as-judge

Multi-agent evaluation pipeline with three specialized AI agents (Researcher, Analyzer, Judge) that score documents against benchmarks. Runs fully local via Ollama.

Code

Agentic Context Engineering

The ACE loop: a live chat system that builds a JSONL Playbook with FAISS retrieval and self-reflection, growing smarter with every interaction.

Code

ParticleFilter

Sequential Monte Carlo particle filter for predicting remaining structural life using the Paris–Erdogan fatigue law. Applied to structural health monitoring.

Code

Deep-SHM

Applying deep neural networks to structural health monitoring — a domain previously limited by sparse labeled data. Explores transfer learning and augmentation.

Code

Writing

I write about language models, low-resource NLP, and applied machine learning on my blog.