Projects | Mohit Tater

Enterprise Ops Workspace

How the app is put together

Three connected modules

Client capture wizard, template library admin, and reporting. Each module has its own role but shares a common data model, so a change to a template ripples cleanly into how it's used in onboarding and how it shows up in reports.

Data model

Clients, templates, submissions, and audit records. Designed so historical onboarding records stay intact and readable even as templates evolve over time — the kind of detail that only matters once you're months into using it.

Auth and deployment

NextAuth v5 against the company's OIDC server for SSO, Supabase service-role on the server side for trusted operations, and a CodeBuild → ECS pipeline for deploys.

AI Resizing Studio

How the three interfaces work together

Processing — generate at scale

Drop in images or a CSV and the pipeline fans each asset out into all of its target ratios at once. Every output gets a risk score (none → high) computed from crop loss, aspect-ratio mismatch, and the strategy used — so problem outputs are flagged before a human ever opens them. Smart / AI-only / Force-AI modes trade speed against cost per run.

Review — a keyboard-fast human pass

Source and generated sit side by side with the AI's verdict and score already attached. Reviewers rate Good / Okay / Bad entirely from the keyboard and move through the whole queue without reaching for the mouse — filters rebuild the queue on the fly, so you can burn down just the unreviewed, or just the high-risk outputs.

Evaluation — AI in the loop

An AI visual judge grades every output against the same rubric a human would. Deterministic checks and a retry/escalation policy then decide what to auto-accept, retry with a stronger safe-area prompt, reroute to another model, or send up for manual design work. The dashboard tracks good-rate, bad-rate on high-risk outputs, and cost per good output — and surfaces where the AI and humans disagree, so the rubric keeps sharpening over time.

Designing Dashboards

What's inside the case study

Six grounding principles

Start from the decision not the data, lead with one thing, earn every element, show the insight not just the data, end in a next step, and spend emphasis sparingly — the questions I designed against from the first pass, transferable to any dashboard.

A real redesign, including the wrong turns

Original → V1 → V4 → V7, shown honestly. It overshoots into decoration before settling — the V4 frame is kept in on purpose, because an iteration that mistook richness for clarity is the clearest lesson of the lot.

Validated, not just asserted

To keep it honest, a fresh agent with no memory of the previous version rebuilds the dashboard blind from the same brief. If it lands in the same place, the principles are doing the work — not attachment to a layout.

Slicely

What's inside

Annotation module

A focused interface for marking up source PDFs — the part you spend the most time in if you're configuring a new slicer. Designed so the heavy work of pointing at the right regions of a PDF feels lightweight.

Dynamic response renderer

Each slicer can return very different shapes of data. The renderer picks the right view automatically — a single number, a chart, a table, or a formatted text block — so the dashboard reads cleanly across slicers without per-slicer UI work.

Bulk processing & webhooks

Designed to chew through batches and notify other systems when results are ready, so it fits into a larger pipeline rather than being a standalone tool.

AI Toolkit

Skills in the toolkit

product-ui

Production-grade product interface work — the invisible craft that makes software feel premium.

dashboard

Modern, actionable dashboards through a structured requirements → strategy → design workflow.

interface-grader

Quantitative grading of any frontend interface against a craft rubric, with iteration-over-iteration deltas.

ux-patterns & interaction-design

Frictionless UX patterns and the microinteractions, motion, and state-handling that go on top.

ui-design-system

React + Tailwind + Radix + shadcn/ui systems — accessible components, theming, dark mode.

harvest-feed

Mine a working session for publishable posts and turn quiet wins into content.

Codex automations

Templates for daily digests, weekly journals, content mining, and session-log backups.

ReviewHub

How the loop closes

Live prototype review

An Agentation widget overlays the running app so anyone — PM, designer, client — can drop pixel-accurate annotations without leaving the browser. Feedback is captured against what the reviewer is actually seeing, not a static screenshot.

Feedback, where your team already lives

Slack is the default destination, but Confluence is now wired in for teams that work out of Atlassian. Pick the destination per project; the rest of the pipeline doesn't change.

Agent handoff and status sync

Coding agents pick tasks up from the feed and execute. Once a fix lands, the Confluence page or Slack thread updates itself — the loop closes without anyone playing telephone between QA, design, and engineering.