MSP Raja ยท Writing https://www.mspraja.com/writing Essays, research notes and a daily lessons series on agents, evaluation and AI in supply chains. en An eval harness is a product, not a script https://www.mspraja.com/writing/eval-harness-is-a-product https://www.mspraja.com/writing/eval-harness-is-a-product Thu, 24 Sep 2026 00:00:00 GMT Evals What changed when our agent evaluations got owners, versions and a release process. Day 212: Retries are a policy, not a loop https://www.mspraja.com/writing/day-212-retries-are-a-policy https://www.mspraja.com/writing/day-212-retries-are-a-policy Fri, 31 Jul 2026 00:00:00 GMT Agentic AI Decide when an agent may retry before it runs, and write the decision down. Ticket to pull request: where coding harnesses fail https://www.mspraja.com/writing/ticket-to-pull-request-where-harnesses-fail https://www.mspraja.com/writing/ticket-to-pull-request-where-harnesses-fail Thu, 30 Jul 2026 00:00:00 GMT Agentic AI Six failure modes from 180 tickets, and the check that catches each one. Planning under semantic constraints https://www.mspraja.com/writing/planning-under-semantic-constraints https://www.mspraja.com/writing/planning-under-semantic-constraints Thu, 18 Jun 2026 00:00:00 GMT Research notes Day 140: Tool descriptions are prompts https://www.mspraja.com/writing/day-140-tool-descriptions-are-prompts https://www.mspraja.com/writing/day-140-tool-descriptions-are-prompts Wed, 20 May 2026 00:00:00 GMT LLMs Exceptions are the supply chain AI problem https://www.mspraja.com/writing/exceptions-are-the-supply-chain-ai-problem https://www.mspraja.com/writing/exceptions-are-the-supply-chain-ai-problem Wed, 06 May 2026 00:00:00 GMT Supply Chain AI Context budgets for multi-agent systems https://www.mspraja.com/writing/context-budgets-for-multi-agent-systems https://www.mspraja.com/writing/context-budgets-for-multi-agent-systems Thu, 12 Mar 2026 00:00:00 GMT Agentic AI Day 35: A grader you cannot read is a grader you cannot trust https://www.mspraja.com/writing/day-35-a-grader-you-cannot-read https://www.mspraja.com/writing/day-35-a-grader-you-cannot-read Wed, 04 Feb 2026 00:00:00 GMT Evals Day 1: One lesson a day for a year https://www.mspraja.com/writing/day-1-one-lesson-a-day https://www.mspraja.com/writing/day-1-one-lesson-a-day Thu, 01 Jan 2026 00:00:00 GMT Research notes Defending AI Systems Against Adversarial Attacks: A Practical Implementation Guide https://www.mspraja.com/writing/defending-ai-against-adversarial-attacks https://www.mspraja.com/writing/defending-ai-against-adversarial-attacks Sat, 20 Dec 2025 00:00:00 GMT Cybersecurity Real-world code and architecture for protecting production ML models from data poisoning, model evasion, and prompt injection attacks with threat analysis and layered defense strategies. Toolformer: Teaching Language Models to Use Tools - Research Annotation & Practical Implementation https://www.mspraja.com/writing/toolformer-research-annotation https://www.mspraja.com/writing/toolformer-research-annotation Fri, 19 Dec 2025 00:00:00 GMT LLMs Deep dive into Meta AI's groundbreaking Toolformer paper showing how LLMs learn to use external tools like calculators, search engines, and APIs. Includes architecture diagrams, implementation guide, and real-world applications in enterprise systems. Voice-Enabled AI Agents in Logistics Operations: Architecture & Implementation https://www.mspraja.com/writing/voice-ai-agents-in-logistics https://www.mspraja.com/writing/voice-ai-agents-in-logistics Fri, 19 Dec 2025 00:00:00 GMT Supply Chain AI Complete technical guide to implementing hands-free voice interfaces for warehouse management and transportation coordination. Includes acoustic modeling for noisy environments, dialogue management architecture, and real-world deployment results achieving 45% faster response times. Advanced IAM Threat Detection Using Graph Neural Networks: Complete Architecture https://www.mspraja.com/writing/iam-threat-detection-with-gnns https://www.mspraja.com/writing/iam-threat-detection-with-gnns Thu, 18 Dec 2025 00:00:00 GMT Cybersecurity Deep technical dive into implementing GNN-based threat detection for enterprise IAM systems. Covers graph construction from identity logs, Graph Attention Networks architecture, temporal embeddings, MITRE ATT&CK mapping, and production deployment achieving 92% detection accuracy with 88% reduction in false positives. What 25 production skills taught me about agent interfaces https://www.mspraja.com/writing/what-25-production-skills-taught-me https://www.mspraja.com/writing/what-25-production-skills-taught-me Tue, 09 Dec 2025 00:00:00 GMT Agentic AI Generative UI needs a contract https://www.mspraja.com/writing/generative-ui-needs-a-contract https://www.mspraja.com/writing/generative-ui-needs-a-contract Fri, 14 Nov 2025 00:00:00 GMT LLMs Shipment data is a graph, not a table https://www.mspraja.com/writing/shipment-data-is-a-graph https://www.mspraja.com/writing/shipment-data-is-a-graph Tue, 28 Oct 2025 00:00:00 GMT Supply Chain AI