Decision Models
Wrote up Typesafe's Jev, a model that returns typed decisions with calibrated probabilities instead of prose. What it claims, how it's built, and what I'd still want to test before trusting it.
Read the field note →AI systems. Security. Things I've actually built.
Notes from the field on agentic AI, cyber security and emerging technology.
Live Lab
Illustrative sketches of the patterns I work with, not production code.
INPUT→PLANNER→AGENTS→VERIFY→ACTION
# Multi-agent orchestration with state management class AgentOrchestrator: async def coordinate(self, task): plan = await self.planner.decompose(task) results = await self.execute_parallel(plan) return self.synthesize(results)# Real-time voice processing and response generation class VoiceRuntime: async def process_audio(self, stream): transcript = await self.transcribe(stream) response = await self.generate(transcript) return await self.synthesize_speech(response)# Decision model evaluation and verification class DecisionModel: def evaluate(self, context, options): scored = self.score_options(context, options) return self.verify_constraints(scored)# Local-first agent runtime without cloud dependencies class LocalAgent: def __init__(self, model_path): self.model = self.load_local_model(model_path) self.memory = LocalMemoryStore()# Security validation for AI system inputs and outputs class AISecurityValidator: def validate_input(self, user_input): return self.check_injection_patterns(user_input) def verify_output(self, ai_response): return self.scan_for_leakage(ai_response)Field Work
Things I've built, tested and learned from.
Recent Field Notes
- Field Note #1: Jev — typed decisions instead of proseNotes from the field on a new model that challenges the basic assumption that AI intelligence has to be expressed as text