https://ahmad.design/writing/ Essays on AI-native product design, auditable AI research, and shipping research as code. Written by Ahmad Bilal, Product Design Engineer & Researcher. en Sun, 23 Aug 2026 23:04:39 +0000 https://ahmad.design/og.pnghttps://ahmad.design/writing/ https://ahmad.design/writing/designing-ai-for-review.html https://ahmad.design/writing/designing-ai-for-review.html Sun, 23 Aug 2026 00:00:00 +0000 Ahmad Bilal When a model can draft anything, the expensive part is the review. What an interface has to expose so a person can disprove the output fast. https://ahmad.design/writing/tb-chest-xray-dataset-bias.html https://ahmad.design/writing/tb-chest-xray-dataset-bias.html Wed, 19 Aug 2026 00:00:00 +0000 Ahmad Bilal My TB screener scored 92 on public benchmarks and 78 at a real hospital. The audit found the benchmarks partly reward recognizing the pipeline rather than the disease. https://ahmad.design/writing/pakistan-sign-language-translation.html https://ahmad.design/writing/pakistan-sign-language-translation.html Wed, 19 Aug 2026 00:00:00 +0000 Ahmad Bilal How I built a Pakistan Sign Language translation system: the context-free-grammar decision, the six-stage pipeline, and what peer review left unsolved. https://ahmad.design/writing/jtbd-for-llm-product-features.html https://ahmad.design/writing/jtbd-for-llm-product-features.html Wed, 19 Aug 2026 00:00:00 +0000 Ahmad Bilal A capability is not a job. How to write Jobs-to-be-Done statements for LLM features, with a worked AutoLeap example and what JTBD couldn't answer. https://ahmad.design/writing/grounded-personas.html https://ahmad.design/writing/grounded-personas.html Wed, 19 Aug 2026 00:00:00 +0000 Ahmad Bilal What the 2022 to 2026 research actually shows about LLM persona generation: what holds, what breaks, and how to ground a persona in evidence. https://ahmad.design/writing/simulated-user-research-workflow.html https://ahmad.design/writing/simulated-user-research-workflow.html Sun, 16 Aug 2026 00:00:00 +0000 Ahmad Bilal How a simulated user research workflow runs: grounded personas, blind interviews, evidence chains, confidence scores, and a thirty-minute audit. https://ahmad.design/writing/pvara-plain-language-guide.html https://ahmad.design/writing/pvara-plain-language-guide.html Sun, 16 Aug 2026 00:00:00 +0000 Ahmad Bilal What PVARA licenses, how NOC applications work today, and where Pakistan's virtual asset law stands, sourced and dated as of August 2026. https://ahmad.design/writing/designing-a-regulator-people-can-trust.html https://ahmad.design/writing/designing-a-regulator-people-can-trust.html Sun, 16 Aug 2026 00:00:00 +0000 Ahmad Bilal Foundation design for the Pakistan Virtual Assets Regulatory Authority website, benchmarked against VARA UAE and ADGM, and what earns a regulator trust. https://ahmad.design/writing/ai-native-product-design.html https://ahmad.design/writing/ai-native-product-design.html Sun, 16 Aug 2026 00:00:00 +0000 Ahmad Bilal AI-native isn't a chat box bolted onto an old product. It's the harness, evals, guardrails, approval gates, made part of the interface itself. https://ahmad.design/writing/agent-experience-design.html https://ahmad.design/writing/agent-experience-design.html Sun, 16 Aug 2026 00:00:00 +0000 Ahmad Bilal What changes when an AI agent reads your product alongside a human: affordances, error states, provenance. A checklist, not a platform pitch. https://ahmad.design/writing/harness-engineering-for-research-agents.html https://ahmad.design/writing/harness-engineering-for-research-agents.html Sat, 15 Aug 2026 00:00:00 +0000 Ahmad Bilal The agent-harness discourse is all about coding agents. Research agents fail differently, and need a different harness. Here is what that harness holds. https://ahmad.design/writing/less-expertise-more-coverage.html https://ahmad.design/writing/less-expertise-more-coverage.html Fri, 29 May 2026 00:00:00 +0000 Ahmad Bilal Casting a model as a domain expert narrows what it treats as in-scope. Run the comparison on your own task: what to measure, and what would prove it wrong. https://ahmad.design/writing/designing-ai-behavior.html https://ahmad.design/writing/designing-ai-behavior.html Fri, 29 May 2026 00:00:00 +0000 Ahmad Bilal I shipped a one-click report button, then killed it. Why the checkpoints came back, what the reversal cost, and why unanimity caps confidence at 75%. https://ahmad.design/writing/building-with-llms.html https://ahmad.design/writing/building-with-llms.html Fri, 29 May 2026 00:00:00 +0000 Ahmad Bilal The rubric came before the prompt and cost about two weeks of demo time. What the checks are, what they caught across 46 routed studies, what they miss. https://ahmad.design/writing/research-to-product-to-code.html https://ahmad.design/writing/research-to-product-to-code.html Thu, 28 May 2026 00:00:00 +0000 Ahmad Bilal A decision rule: collapse the research, design and engineering relay when the definition of done keeps moving, and keep the handoffs when it holds. https://ahmad.design/writing/grounded-simulation.html https://ahmad.design/writing/grounded-simulation.html Thu, 28 May 2026 00:00:00 +0000 Ahmad Bilal An architecture for LLM-based synthetic user research: hypothesis-blind personas, grounding in behavioural science, and no finding that skips its evidence. https://ahmad.design/writing/auditable-ai-research.html https://ahmad.design/writing/auditable-ai-research.html Thu, 28 May 2026 00:00:00 +0000 Ahmad Bilal A trace is not a truth. How to audit an AI-generated finding in thirty minutes, and why a clean evidence chain still never proves a claim is valid. https://ahmad.design/writing/ai-as-a-design-material.html https://ahmad.design/writing/ai-as-a-design-material.html Thu, 28 May 2026 00:00:00 +0000 Ahmad Bilal A model fails plausibly, so build the approval gate first: where a person signs off before anything commits. One shipped artifact, walked end to end.