AI Nerd
Built Marlo, the agent learning platform. Capture behavior, evaluate outcomes, and turn failures into learnings that deploy back into the agent. Intelligence compounds over time.
- Marlo SDK -- Python & TypeScript SDKs for capturing agent traces, tool calls, and reasoning
- Marlo Inbox -- AI email & calendar assistant powered by Marlo
- Marlo Docs -- Documentation for the platform
Continual Learning, Not Training: Online Adaptation For Agents
ATLAS achieves 54.1% success with GPT-5-mini as Student, outperforming GPT-5 (High) by 13% while reducing cost by 86%. Cross-validation improves accuracy from 28% to 41% without retraining.
Built the open-source continual learning stack for production LLM agents.
- ATLAS -- Continual learning framework. Teacher-Student architecture with cross-domain transfer and inference-time adaptation. +15.7% accuracy, +31% task completion, 97% non-degradation rate.
- Atlas SDK -- Runtime wrapper capturing causality traces. Enables offline training via GRPO and GKD.
- CL-Bench -- Benchmark for evaluating agent continual learning in stateful environments. Production-realistic CRM workflows with multi-turn conversations and state mutations.
- HuggingFace Models -- Open weights for continual learning research.
- OpenSearch -- AI research engine that searches, analyzes, and generates cited reports with charts(Better Perplexity)
- ReplicaDB -- Single platform for data and analytics needs
- CI Code Companion -- AI-powered GitLab CI/CD tool using Vertex AI for auto-generating tests, docs, and pipeline templates
- Continual learning for autonomous agents
- Reinforcement learning from production feedback
- Multi Agentic Systems
- AI -> AI interaction
- Cross Domain transfer in AI
- Agent reliability at scale
- Email: amanwork2025@gmail.com
- Company: Rockland System
- Research: arxiv.org/abs/2511.01093