DECISION ARCHIVE
Every call I made, and what it cost.
A product manager is the sum of their trade-offs. This is the full log across four programmes — filterable, and each one links back to the case it came from.
Tax Assistant ChatbotRefuse rather than answer when retrieval is weakA hedged wrong answer about tax still gets acted on. Refusal routes the user to a human in one step instead of three.COSTRoughly 18% of sessions end in escalation rather than resolution.Tax Assistant ChatbotCite the circular on every answerThe citation is what makes the answer checkable by an accountant later. It also disciplined the corpus — uncitable answers exposed gaps.COSTLonger, denser replies that test worse on first impression.Tax Assistant ChatbotPersian-only at launchThe evaluation set only existed in Persian, and shipping an unevaluated English path would have doubled the surface we could not vouch for.COSTNo English-speaking users at launch.Nutrition AssistantShow 'unreadable' rather than a low-confidence verdictA wrong 'safe' has a medical cost. Unreadable is an honest state that the user can act on by reading the label themselves.COST14% of scans return nothing useful.Nutrition AssistantDrop food logging from onboardingLogging was the reason users churned in week one. The scan is the job; the diary was our idea, not theirs.COSTLost the longitudinal data set the analytics roadmap assumed.Nutrition AssistantThree states, no numeric scoreA score invites interpretation the model cannot support. Three states are checkable and unambiguous at a glance.COSTLess differentiated in a store listing full of scored competitors.NLP Comment AnalysisReport movement, not absolute sentimentAbsolute sentiment across platforms is not comparable — review cultures differ. Movement within one source is a real signal.COSTNo headline ranking, which is what leadership initially asked for.NLP Comment AnalysisHuman-defined taxonomy before modellingClusters that don't map to owned feature areas produce insight nobody can action. The taxonomy made every finding assignable.COSTThree weeks up front and a blind spot for genuinely novel themes.NLP Comment AnalysisOutput questions, not conclusionsThe pipeline sees complaints, not causes. Framing output as questions kept the interpretation with the people who had context.COSTLower perceived sophistication in demos.AI Customer ServiceAutomate by decision class, not by queueVolume and error cost are not correlated. Classing the work first let us automate 60% of items and leave the expensive 40% untouched.COSTA slower headline number in the first quarter.AI Customer ServiceRules layer holds the boundary; the model drafts inside itA threshold is a statistical promise. Auditors need a deterministic one — the rules layer is what we can defend in a review.COSTExtra engineering, and some correct model calls get blocked.AI Customer ServiceTwo queues stay manual, permanentlyError cost in those classes exceeded the salary saved by roughly 4×. Automating them would have been a net loss dressed as progress.COSTThe programme's savings headline is smaller.AI Customer ServiceShip the audit trail before the first automated decisionReconstruction after the fact is not reconstruction. Every automated decision has carried its full input snapshot since day one.COSTSix weeks before any automation went live.Traffic Analysis DashboardShip a focused operational toolProfessional traffic analysis dashboard built with Streamlit for Iranian highway data with ML predictions and holiday integration.COSTNarrower initial surface, faster time to value.Guilds Complaint DashboardShip a focused operational toolAn intelligent dashboard for managing and analyzing complaints from guilds and trade associations with real-time monitoring.COSTNarrower initial surface, faster time to value.HR Management SystemShip a focused operational toolComprehensive document management system built with Laravel for HR database with employee records and multi-language support.COSTNarrower initial surface, faster time to value.Smart Multi-Carrier RoutingShip a focused operational toolAdvanced routing system with OR-Tools optimization algorithm for package distribution and optimal route calculation.COSTNarrower initial surface, faster time to value.Nova AI OptimizationShip a focused operational toolTechnical project using genetic algorithms (MVRP) for logistics optimization, reducing costs by 25%.COSTNarrower initial surface, faster time to value.Atlas 3D Bin PackingShip a focused operational toolML tool for optimizing cargo loading, increasing vehicle utilization by 30% through 3D visualization.COSTNarrower initial surface, faster time to value.