Inspiration

By 2030, one in six people on Earth will be over sixty, and around four in ten older adults already take five or more medicines, often prescribed by different doctors who never talk to each other.
While reading about medication safety, I came across the prescribing cascade: a side effect of one medicine is mistaken for a new illness, so another medicine is added to treat it, and then another to treat the side effects of that one.
The classic example became our example patient, Mdm Tan, a 78-year-old woman who lives alone. A neighbour gives her ibuprofen for her knee, ibuprofen raises her blood pressure, her doctor starts amlodipine, amlodipine swells her ankles, and the usual next step is a water pill.
A 2025 study of over half a million older adults found that this exact chain was the strongest cascade of all.
The clues are almost always there, in what people say between doctor visits, such as "my ankles are swollen again." Her doctor never hears about the ibuprofen, her daughter never sees when the swelling started, and she only knows that she feels worse.
Nobody is listening for the pattern, and I wanted to build something that does.

What it does

EchoCare is an AI medication-safety companion for older adults who live alone, and for the families and clinicians who look after them.
Her app shows one thing at a time, in very large text, with pictures of her actual pills. Reminders chime, flash and read themselves aloud.
She can tell EchoCare how she feels by tapping an answer, typing, or simply speaking in her own words.
EchoCare compares every symptom with when each medicine started, using 25 published prescribing-cascade rules from ThinkCascades and the AGS Beers Criteria, and checks for dangerous interactions and duplicate ingredients.
When it finds a likely cascade, it never tells anyone to stop a medicine. It turns the finding into a clear question for her doctor, because the doctor should always decide.
Her daughter sees everything live on her own phone through a shared care circle, with the most important finding shown first in plain words.
If a dose is missed for two hours, or she presses "I need help", her daughter is alerted within seconds.
If her daughter tries to add a water pill for the swollen ankles, EchoCare warns her before anything is saved that the swelling may come from amlodipine.
Before each appointment, EchoCare writes a one-page memo for the doctor covering her symptoms, her medicines from every source, missed doses and each possible cascade with its evidence, printable as a PDF.
The memo saves consultation time: instead of piecing together the patient's history during a short visit, the doctor grasps the full context at a glance, leaving more time for the decisions and conversations that matter.
Pharmacists see their patients sorted by medication risk, so the people most at risk are reviewed first.
The whole app switches between English, Mandarin, Malay and Tamil with one setting.
Emergency phrases such as chest pain are recognised offline in all four languages, so she is told to call for help and her family is alerted even when the AI is unavailable.

How we built it

EchoCare is a plain HTML, CSS and JavaScript web app with no framework and no build step, so it runs on almost any phone.
The demo website is hosted on Cloudflare Pages, and a small Python script bundles the same files into a single page for the full version on claude.ai.
Claude reads her everyday words and photos of medicine boxes, and turns them into structured records of symptoms, doses, new medicines and mood.
The safety decisions are deliberately kept away from the AI and come from a rule engine instead. Each rule links a first medicine $A$ to a known side effect $s$ and a follow-on medicine $B$ often prescribed to treat it.
For every rule, the engine gathers independent kinds of evidence, where $t_A$ and $t_B$ are the dates each medicine started and $t_s$ is when the symptom began:

$$ \begin{aligned} \text{sequence:}\quad & 0 < t_B - t_A \le 180 \text{ days} \ \text{symptom:}\quad & t_s - t_A \ge -3 \text{ days, and the symptom is unresolved} \ \text{report:}\quad & \text{her family reported } s \text{ as a suspected side effect of } A \end{aligned} $$

Because the symptom counts as evidence on its own, a finding can appear before $B$ is ever prescribed, which is how EchoCare can stop the third prescription.
To keep false alarms down, each finding is graded by $k$, the number of distinct kinds of evidence, and pairs usually prescribed together on purpose never rise above possible:

$$ \text{rating} = \begin{cases} \text{Likely}, & k \ge 2 \text{ and the pair is not usually intended} \ \text{Possible}, & \text{otherwise} \end{cases} $$

US FDA adverse-event data for 89 common medicines, filtered to people over 65, is built in as context for the doctor but never raises a rating on its own.
A shared care circle keeps one live record in sync between her phone and her family's phones.
jsPDF generates the printable doctor memo.
A translation layer swaps the text of every screen after it is drawn, treating names, numbers, medicine names and dates as placeholders so one translation works for any patient.
The app runs its own engine check, with textbook cascade cases and false-alarm controls, every time it is opened.

Challenges we ran into

Catching cascades without crying wolf: getting the timeline logic, the evidence grading and usually intentional combinations, such as a laxative alongside an opioid, right took many rounds of testing against control cases.
Separating the roles of AI and rules: it was tempting to let the AI decide everything, and choosing to have the AI listen while published rules decide shaped the whole design.
Handling emergencies without the AI: "dada saya sakit" in Malay, or chest pain in Mandarin or Tamil, had to trigger the right response offline, while "no chest pain today" could not raise a false alarm.
Designing for a 78-year-old: every screen had to work with low vision, unsteady hands and hearing loss, so reminders flash and vibrate as well as chime, and each screen asks for only one thing.
Translating the whole app without rewriting it, including fixing an encoding bug that turned Chinese and Tamil text into garbled characters on some pages.
Bugs that only appeared over time: the app fixed the date when the page first loaded, so a phone left on overnight stopped reminding her the next morning, until I made it start each new day on its own.

Accomplishments that we're proud of

The built-in engine check catches all 13 textbook cascades with no false alarms across 6 control cases.
The language check passes all 11 emergency and parsing cases across four languages.
The central moment works from start to finish: she says her ankles are swollen, her daughter's phone shows the whole chain, and EchoCare stops the water pill before it is ever added.
The doctor memo turns weeks of everyday conversations into one page a doctor can read at a glance before the consultation.
Emergencies are handled safely even when the AI is switched off.
It is honest about what it cannot yet do, and offers questions for the doctor rather than instructions.

What I learned

Clinical insight matters more than clever AI: the most valuable part of EchoCare is a simple, well-sourced rule applied at the right moment.
Safety is a design choice, and deciding from the start what the system must never do made every other decision easier.
Accessibility is the product, not a feature: if a 78-year-old cannot use it on a bad day, nothing else matters.
Testing unglamorous cases, such as false-alarm controls, other languages and what happens after midnight, caught real bugs that would have mattered to real people.
AI lowers the barrier to building: with Claude Code as a partner, I was able to build something this ambitious on my own.

What's next for EchoCare

A pharmacist review of every rule, then a pilot with a clinic or pharmacy that measures how often clinicians agree with EchoCare's flags.
Native-speaker review of the Mandarin, Malay and Tamil translations, and more languages to reach older adults around the world.
Push alerts that reach a locked phone.
Phone-call check-ins for older adults who do not use a smartphone.
Better voice input that copes with pauses, accents and mixed languages.
A carer-led mode for people with advanced dementia, whose own reports become less reliable over time.