Fynd · Agentic Impetus Product Design & Front End
From a business exception
to a considered action.
A retail workspace to inspect AI reasoning and turn business exceptions into owned work.
Explore the key decisionThe starting point
Seeing a problem is only the start.
Reports and approvals spanned Excel, email and calls. A leader needed to understand an exception, judge the recommendation and assign follow-up work—not just read another dashboard.
I connected evidence, recommendations and task ownership in one workspace, then built the front end and integrated APIs with Claude Code and Codex.
- The difficulty
- Find an issue in a report; carry its context into a separate conversation and follow-up.
- What I changed
- Inspect the recommendation beside its evidence; open the linked task to see its owner and status.
01 / Investigate
Keep the recommendation
beside the evidence.
A recommendation needs a reason. I kept analysis separate from suggested action, with the exception queue alongside so investigation does not lose its place.
Less space per panel; more context. Leaders investigate one issue without losing the queue.
02 / Follow through
Connect the insight
to the work it creates.
The agent can connect insight → exception → linked task automatically. Business users can trace each task back to its reason.
- 01 / Detect
Insight
The agent identifies a business signal and explains what it means.
- 02 / Surface
Exception
A finding that needs attention becomes an Action Register entry.
- 03 / Initiate
Linked task
The agent can create the task automatically and link it to the exception.
The screens show the manual task-creation path: edit the brief, assign an owner and follow its status. The source exception stays linked; creating a task does not mean resolving the issue.
03 / Design through delivery
Keep the briefing.
Change the composition.
On mobile, I led with the briefing and urgent exceptions, then put individual findings one level deeper. Bottom navigation keeps the main areas available without compressing the desktop layout.
I used Claude Code and Codex to implement the React and TypeScript front end and integrate APIs, refining loading, error and responsive states in the running product. Category and format views reused the existing CEO architecture.
Business standards and AI workflows
Brain / Specs & workflows
Make the standard inspectable.
One developer and I built the first version of Brain using Hermes Agent over a weekend, then continued refining the feature afterward.
I organised business responsibilities first, with SOPs, skills and reviewable criteria underneath. People can find their work without knowing an agent’s name.
The wider Impetus context
The work primarily served Reliance Retail’s Fashion & Lifestyle and Grocery teams, moving from spreadsheet-and-email workflows towards a shared platform.
The broader Impetus brief targeted a fast-fashion model for F&L, with Design to Shelf in 45 days as a transformation goal—not an outcome attributed to this interface.
Impact
Less escalation. Earlier action.
- fewer CXO escalations
- 20%
- 100 → 80 escalations over two months
- Routine exceptions become assigned tasks; critical decisions reach CXOs with the evidence and context needed to act.
- less aged inventory
- 12%
- Representing ₹4.34 Cr in stock value
- Threshold-based monitoring and repeatable workflows help teams respond to demand and stock risks. The stock value represents the same inventory reduction, not cash savings.
Inspect before acting
Evidence and advice remain distinct. Leaders can review the reasoning behind an exception before deciding what to do.
Follow through with context
Source, owner and status remain connected as an exception becomes a task. The design makes responsibility traceable rather than leaving follow-up outside the workspace.
What I would validate next
Next validation step
Traceability of the decision
- Representative task
- Follow an exception into a task, identify its owner and explain the evidence behind it.
- What to measure
- Count tasks with a traceable source, owner and next action. Ask leaders to explain why they accepted or rejected the recommendation.
Pair completion with evidence comprehension; accepting more AI recommendations is not automatically a better outcome.