Applied AI research.
Production lessons.
Experiments, technical analysis, and implementation lessons from Algen's product and applied research team.
All notes
13 posts
Approvals Belong in the Runtime, Not the Prompt
“Always ask before sending an email” is a suggestion, not a control. Four ways prompt-based approvals fail, and what an approval in the execution path needs.
AI EngineeringSaurav Samantray7 minWhy We Built AgentStripes: A Trust Rating for AI Agents
Agents now refund orders and change live records, but there's no easy way to tell if one is safe to use. So we built AgentStripes, a 0 to 7 trust rating.
AI Agent Maturity ModelVijay Poudel3 minWhere Should Agent Code Live? Runtime, Tools and Agents at Algen
Most agent code isn't specific to any one agent. One question, “who else would need this?”, decides whether it belongs in the runtime, in a connector, or in the agent itself.
AI EngineeringSaurav Samantray9 minHow to Evaluate an AI Agent Before You Deploy It
Evaluate AI agents before deployment with a 40-capability checklist across 12 areas, graded by evidence. Rate yours in 30 seconds with AgentStripes.
AI GovernanceAlgen.AI Research9 minSmall Models, Narrow Jobs: What Fine-Tuning SLMs for a Talk-to-Data Agent Taught Us
Row counts lie, validation loss lies, and “I don't know” beats a confident wrong answer. Lessons from replacing GPT with small fine-tuned models.
AI EngineeringSaurav Samantray9 minJev Explained: How TypeSafe’s Decision Model Fits Into AI Agents
A practical look at typed decisions, calibrated probabilities, agent architecture, and the evidence behind Jev’s performance claims.
AI EngineeringVijay Poudel12 minReasoning vs Standard LLMs: When Is It Worth Paying the Thinking Tax?
A local side-by-side benchmark exploring when reasoning models pay off versus standard LLMs — and how to allocate thinking time to task risk.
EngineeringSaurav Samantray5 minCode Is Cheap, Context Is Expensive: The Strategic Power of the Forward Deployed Engineer
95% of enterprise AI pilots fail not because of model quality, but because nobody is sitting close enough to the problem to see what the model actually needs.
EngineeringSaurav Samantray7 min