Applied AI · A study in human control
A suggestion.
Not a verdict.
The interesting part starts after the model answers. Can someone inspect the recommendation, change it, save it—or choose a different path entirely?
The answer is the beginning of the interface.
Early experiments with models, embeddings, agents and recommendations made generation look like the central event. In a complete product, it is one step among many. A plan still needs to fit the available time. A suggestion needs an identifiable place in the data. A saved choice needs to survive the next visit.
Planning tools and specialist assistants moved my attention toward that surrounding work. I connected generated content to structured records, discovery, persistence and editable journeys. The model could offer a possibility; the product had to make that possibility usable.
This means defining what output is allowed to contain before deciding how elegantly to present it. Stable domain structure belongs in a schema. Descriptive content can remain flexible within that boundary. Validation is the point where a plausible sentence becomes something the application can safely handle.
An ambitious afternoon
- Visit the gallery
- Cross town for a second stop
- Fit in a final stop before dinner
A plausible plan can still be too much for the person using it.
Count the effort after the suggestion.
A recommendation can be impressive and still create work. If changing one stop means regenerating an entire plan, the person is negotiating with the tool instead of using it. If a rejected suggestion returns unchanged, the interface has not really listened.
I wanted editing, confirmation and rejection to remain visible. The useful comparison is not just the time taken to generate something. It is the effort needed to arrive at an acceptable decision, including inspection and correction.
Richer context can improve relevance, but it also brings cost, latency and privacy considerations. More model freedom can produce novel possibilities while making validation harder. These are product choices. They deserve the same attention as the model selection.
The person should be able to disagree—and still finish the task.
Keep a door open when generation fails.
Manual discovery was an important part of the planning work. It kept the product useful when a recommendation was unavailable, irrelevant or simply not what someone wanted. The conventional path was a real route through the application, with its own complete states.
That changes the role of intelligence. It becomes a collaborator inside an understandable workflow rather than the only way into the product. People can accept its help without surrendering the ability to inspect or reverse a choice.
A saved recommendation also needs to be distinguishable from a fresh generation. If someone returns to a plan, silently replacing it with a new answer changes the meaning of “saved.” Persistence should protect the person’s decisions, including their edits, rather than treat every visit as another opportunity to regenerate.
It helps to name the moments separately: suggested, reviewed, edited, confirmed. They do not all need badges in the interface, but the product needs to know the difference. The transition into a committed decision is where a conversational response becomes application state.
I remain interested in what these systems can make easier. The lasting design question is where responsibility sits when a suggestion becomes an action. An exit, an edit and a clear confirmation are small interface details with considerable weight.
Related work & references
Selected public entry points and related portfolio pages. A company website or profile is context, not a claim that private source code is available.
Scope, context & further questions
Supported by public machine-learning, planning, recommendation, bot, storage, and domain-assistant application work.
- Across the whole application: End-to-end. Planning APIs, web experiences, bots, storage, and landing journeys formed complete product flows.
- Data strategy: Hybrid. Structured transactional data coexisted with flexible descriptive and generated content.
- Product requirement: Fallback. Manual discovery remained usable when recommendations were unavailable or unsuitable.
Questions I am still working through
How should products communicate why a recommendation appeared?
Show the few factors that materially influenced the recommendation, separate known facts from generated inference, and expose editable inputs. Explanations should help a person evaluate the proposal, not imitate a hidden chain of thought.
The challenge is choosing explanations that are faithful enough to be useful without creating false certainty.
Where must human confirmation remain mandatory even when model confidence is high?
Where actions are difficult to reverse, materially affect money, health, identity, privacy, access, or another person. Confidence does not remove accountability.
Teams still need domain-specific thresholds for what counts as consequential enough to require explicit confirmation.