Products running in production

A multi-tenant SaaS that pharmaceutical companies subscribe to for medical content production — with a review and approval trail built into the core, not bolted on.

The hard part

Pharmaceutical content cannot go live on "the AI said so". The hard part was making review and approval a property of the data model — every generated asset carries who approved it and what it was based on — so the output survives medical-legal scrutiny. That governance layer is the product; the generation is the easy half.

What it does

• Separate studios for medical writing, campaigns, social creatives, video and training
• Multi-tenant: every pharma company is its own isolated account
• Draft → review → approve workflow recorded against every asset
• Generated content is tied to real, verifiable references

Built with

A profit-aware Amazon advertising analyser for Indian sellers — the segment too small for enterprise ad tools and too price-sensitive for a consultant.

The hard part

Statistical honesty. Most ad tools flag a keyword after an arbitrary click threshold, which generates confident advice from noise. This uses a beta-binomial confidence model, so it stays quiet until the data actually supports a call. The free tool also reads .xlsx files with no dependencies at all, decompressing them natively in the browser.

What it does

• Reads seller spreadsheet exports in the browser, with no upload required for the free tool
• Break-even ACOS calculated from real margins, not a generic benchmark
• Recommendations grouped into costed actions the seller can act on
• Subscription tier: client accounts, saved report history, month-on-month comparison
• A decision record — dismiss a suggestion once and it stops coming back

Built with

Turns an approved audio-visual script into a Premiere-ready timeline, with an approval gate that keeps AI video safe for regulated review.

The hard part

Image-to-video instead of text-to-video, on purpose. Approval happens on a frozen frame, so the video model can only add camera motion — it can never invent medical content. A server-side cron queue runs the submissions, because a twenty-one-scene film cannot live inside a PHP session, and an interrupted call is never silently retried since it may already have been charged.

What it does

• Parses a scene-formatted script into individual scenes automatically
• One still image per scene, approved by the client before anything animates
• Approved stills sent to an image-to-video model; narration generated separately
• Exports FCP XML plus media and a shot list — Premiere imports it directly

Built with

An intention-led dating platform: a React single-page app on a hand-written PHP API, installable as an app on a phone.

The hard part

A single ~1,700-line PHP API entry point handles the whole product, including a hand-rolled SMTP mailer, because shared hosting cannot be relied on to deliver local mail. The service worker caches the app shell but is explicitly forbidden from caching the API or member uploads — cache a dating inbox once and you have leaked one member's messages to another.

What it does

• Password, email one-time-code and Google sign-in
• Profile setup, filtered discovery, interests, matches and messaging
• AI-assisted conversation openers and photo feedback
• LinkedIn verification, plus reporting, blocking and a moderation panel
• Progressive web app — installs to the home screen, works offline

Built with

An Indian nutrition platform with AI diet planning and a retrieval-grounded coach that answers from a curated knowledge base.

The hard part

Running an AI coach at a cost that survives contact with real traffic. Retrieval happens first against the knowledge base, the model is given a tight token budget, and adaptive thinking is explicitly disabled — left on, it silently consumed the budget and returned empty answers. The result costs roughly half a rupee per answer, and is rate-limited per session. Clinically, it uses Asian-Indian BMI cut-offs rather than the WHO defaults, because the wrong thresholds give the wrong advice to the audience it serves.

What it does

• Goal wizard that produces a personalised diet plan
• RAG chat coach answering from a vetted nutrition knowledge base
• Food search, intake tracker and a body-age calculator
• Accounts, saved profiles and plans delivered by email
• Installable progressive web app with an offline fallback

Built with

A mobile-first medical reading platform for doctors — articles, case studies, conference highlights, guidelines and algorithms, with a clinical tutor on hand.

The hard part

Doctors read in short gaps, on a phone, often on hospital wi-fi. Everything is therefore structured around getting to the answer quickly — sectioned content, algorithms rather than essays, and a tutor for the question the article did not cover. The reading experience sets the technical budget, not the other way round.

What it does

• Articles, case studies and conference highlights in one reading feed
• Clinical guidelines and treatment algorithms, organised by section
• A clinical tutor that answers a question the moment it comes up
• Accounts and registration, so reading history follows the reader
• Built mobile-first, for reading between patients rather than at a desk

Built with

A members-only photo archive and contest platform for a bird photography community, replacing group chats where photographs were getting buried.

The hard part

Public member profiles live at the site root, so the profile route is a catch-all registered last — and a reserved-word list stops anyone claiming a username that would shadow a real page. Get that wrong and one member signing up could take a section of the site offline. Image handling runs server-side with proper derivative sizes, because a community archive is mostly photographs.

What it does

• Searchable photo archive by species, location and tour
• Per-member species collection with tick marks as they are photographed
• Voting, plus weekly and monthly contests with jury scoring
• Tour albums for the operators who run the trips
• AI-assisted species suggestions from the photograph itself

Built with

A travel publication of well over a hundred articles, with a hand-built homepage that updates itself from the content behind it.

The hard part

Two things that quietly wreck a content site, both fixed: the site title was empty, so every one of roughly a hundred and seventy pages rendered a broken, dangling browser title — repaired in one move at the settings level rather than page by page. And eight articles had images pasted in as base64 data directly inside the article text, the worst being a single page carrying nearly twelve megabytes of uncacheable markup with no lazy loading.

What it does

• Custom homepage: hero photo strip, destinations, categories and author section
• Latest-articles grid that pulls live from the site's own content
• Scoped styling that cannot collide with the underlying theme
• Photo lightbox and live published-post count

Built with

A short-fiction site for quiet psychological horror, running on a dark theme built from scratch for the reading experience.

The hard part

Every story carried its own pasted inline style blocks — dozens of hardcoded light colours and bespoke class names — so any dark theme showed cream boxes stranded inside a dark page. The theme strips those at render time and restyles the class names itself. The database is never modified, which means removing one filter restores the original appearance exactly.

What it does

• Custom theme built from nothing — no page builder, no parent theme
• All-serif typography chosen for long-form reading
• Dark palette designed so the writing sets the mood, not the interface
• Reading-first layout across phone and desktop

Built with

Study notes for PMP candidates, built on a concept-first approach — understand the idea, then answer the question, rather than memorising a question bank.

The hard part

Most exam sites sell volume — thousands of practice questions that teach pattern-matching instead of understanding. This one deliberately narrows: concepts first, in plain language, and it states up front what it will not do for you. Restraint is the product decision.

What it does

• Study notes written from real exam preparation
• A concept-first path: the idea before the exam technique
• Clear scope — it is explicit about what it does and does not cover
• A structured route through the material rather than loose articles
• Guidance on how to actually use the site to prepare

Built with

A continuing medical education platform for healthcare professionals: evidence-based courses with progress that stays visible and completion that means something.

The hard part

Medical education fails on abandonment, not on content. The product is built around resumption — every screen answers "what do I do next" — and completion is tied to genuine progress rather than time on page, because a certificate that means nothing is worse than no certificate.

What it does

• Evidence-based CME courses for healthcare professionals
• Progress tracked and visible, so a part-finished course is easy to resume
• Meaningful completion rather than a click-through certificate
• Accounts with email and Google sign-in, and address verification
• A guided path built around whatever the next action should be

Built with

A question board for a taught course, where students raise queries and teachers answer them in one place instead of losing them across email and chat.

The hard part

Course questions normally scatter across email, chat groups and corridor conversations, so the same thing gets asked repeatedly and no answer is ever findable again. Putting them on one roled board makes the answer reusable. The front end talks to a small PHP API and holds its session as a token, which keeps the interface fast and the server simple.

What it does

• Separate student and teacher roles, chosen at registration
• A shared board where questions are raised, tracked and answered
• Accounts with sign-in, so a question stays attached to the person who asked it
• Built around a running course rather than a generic forum

Built with