RS
RAM SONI
AI Work · 2025 — Present

Practical AI builds, shaped like products

Pilot products with real businesses, each designed and built end to end with AI. The live pilots are in near-daily use, most of them for months (the AI assistant is the newest), each has automated data backups, and new features keep being added from what users ask for. Client names are left out.

How I Approach It

The AI is the accelerator. The differentiator is choosing the right problem, shaping the workflow, and knowing when simple is stronger than clever.

1
EducationIn active use

A teachable AI assistant for an education consultancy

The problem
A small counseling team works across a CRM, a busy inbox, and an assessment platform. Follow-ups, fee flags, and report sending depended on one person remembering everything.
What I designed and built
An AI assistant that works across those existing tools: it checks the inbox twice a day, triages mail, follows the lead pipeline, flags fees due and data problems, and researches when asked.
Key decisions
Ground rules from day one: it never sends an email, deletes anything, changes fees, or contacts a family without the founding counselor's approval. It learns the way a new team member does, through written procedures I teach it, and keeps a learnings log the team can read.
Where AI fits
The assistant is the product: an AI agent working on the team's existing tools, so there is no new system to adopt.
What I learned
A real assistant is a system, not a chatbot: rules, procedures it can be taught, and the habit of asking first. It keeps people in the loop instead of replacing them, and it is the model I am now taking to the hospital work.
2
Salons and other service businessesIn active use

Stock control for service businesses

The problem
Businesses like salons use products to serve every customer, so stock is part of the service. Usage was tracked by hand, reorders were missed, and nobody could see what was being used up.
What I designed and built
A mobile-first stock app for owners and staff, with product records, stock movements, staff access, low-stock alerts, and CSV export.
Key decisions
Designed for a busy service floor: scan first, few taps, plain language, and glanceable stock health instead of warehouse-style screens.
Where AI fits
Inside the product, AI identifies a product from a photo; when it can't, a regular barcode scanner is the fallback. The product itself was built with AI coding agents.
What I learned
Fitting the floor routine mattered more than adding features.
3
Real estateIn active use

Inventory and lead matching for local estate agents

The problem
Local, unorganized estate agents and their customers run on WhatsApp. Property inventory lives in notebooks or, at best, spreadsheets; photos and videos sit somewhere else; and listings have no standard specs, so properties are hard to compare or match to buyers.
What I designed and built
A mobile workspace with a searchable property inventory, standard specs for every listing, photos and videos kept with each property, and lead management with reminders, synced with Google Sheets and Drive.
Key decisions
Kept WhatsApp at the center instead of replacing it, because that is where agents and customers already talk. Every listing is captured with the same fields, such as configuration, area, floor, price, parking, and possession date, so an agent can put two properties side by side and check them against what a buyer asked for.
Where AI fits
AI lead matching, in progress: matching each lead's needs against the inventory so the agent sees the best-fit properties first. The app was built with AI-assisted development.
What I learned
It was also a proof point that a plain Google Sheet, a tool these agents already know, can serve as the database. At the most local level, the win is structure that fits existing habits: WhatsApp stays, the notebook goes.
4
HealthcareIn build

Finance and operations approvals for a hospital

The problem
Salary runs, vendor payments, and statutory deadlines ran on spreadsheets and messages, with approvals and the audit trail spread across people.
What I designed and built
A role-based system with an obligations register and reminders, a vendor master with protected bank details, bulk payment preparation, two-person approval, print packs, and an audit log.
Key decisions
Approvers get a short list on their phone, and each item opens with its basis: what changed, who gets paid, and which checks passed. Vendor bank details are locked: changing them needs one person to approve and another to call the vendor on the number already on record, a standard guard against payment fraud from fake change requests.
Where AI fits
AI agents prepare drafts: reading invoices and contracts, explaining attendance differences, writing a short morning brief. Each job can use a different model with a fallback, staff and patient names are never sent, and every call is logged. Next: an assistant like the education one, working alongside the hospital's existing tools rather than replacing them, with compliance handled first.
What I learned
The real design work is the approval step: showing a person exactly what the AI prepared and why, so they can decide quickly.
5
EducationIn active use

Lead management for an education consultancy

The problem
Student and parent enquiries were spread across calls, phone contacts, and spreadsheets, and follow-ups were easy to miss.
What I designed and built
A CRM with pipeline and list views, search, duplicate handling, message templates, and lead stages, synced with the Google Sheets and Contacts the team already used.
Key decisions
Kept the CRM close to familiar Google tools and added structure around them, so the team gained follow-up discipline without giving up their habits.
Where AI fits
Designed, built, and iterated with AI-assisted development, with each round shaped by the counselors' feedback. It is also the system the AI assistant now works in.
What I learned
Adoption came from fitting existing habits, not replacing them.
6
Design systemsIn active use

A design system for AI coding agents

The problem
When AI agents write interface code, every screen can drift in spacing, color, and pattern unless the system tells them exactly what to do.
What I designed and built
A lightweight UI kit with design tokens, ten contrast-checked themes in light and dark, components, and instructions written for agents as much as for people.
Key decisions
One theme file per client, no build step, and a checker that reports inconsistencies so quality can be verified, not just hoped for.
Where AI fits
Written to be read and followed by AI coding agents, and used in my newer builds.
What I learned
Agents follow a design system well when its rules are explicit and checkable. A design system is now an instruction set.
7
EducationIn active use

Report generation for an assessment practice

The problem
Individual assessment reports were prepared and emailed one at a time by hand.
What I designed and built
A desktop app that generates the reports and sends them by email.
Key decisions
Kept it a local desktop tool so the operator stays in control of each report before it goes out.
Where AI fits
Designed and built with AI-assisted development.
What I learned
Not every useful tool needs to be a web app; the right form follows how the work is actually done.
8
Desktop softwareIn active use

Backup and sync tool for Mac

The problem
Backups to external drives and network storage need to run reliably without anyone remembering to start them.
What I designed and built
A desktop app with scheduled and drive-triggered backups, automatic reconnection to network storage, and pause and resume.
Key decisions
Quiet by default: it recovers on its own and only asks for attention when something genuinely needs it.
Where AI fits
Built with AI coding agents, backed by 170+ automated tests and versioned releases.
What I learned
AI writes code quickly; tests and release discipline are what make it safe to ship to other people.

Day to day I build with Claude Code, Codex, and Gemini, and run AI assistants with Grok Bot. The tools change quickly; what carries over is the judgment about where AI belongs and how to ship it responsibly.

Start a conversation