App Atlas

AuraHealth is a free health check-in that lives on your phone. It reads your own lab reports, keeps records for you and the people you care for, and takes real measurements with the phone's camera, microphone and motion sensors: fingertip and face pulse, heart-rate variability, breathing, voice, hand tremor, walking, balance, reaction time, hearing and near vision. Every scan checks it is measuring the right thing before it reports a number.

A team of agents on the phone turns this into help: a weekly coaching plan built from your own numbers, medicine checks against your lab results, food verdicts from a barcode or label, symptom signposting to 999 and 111, appointments and re-test reminders, and a summary to take to your doctor. A safety agent reviews every message, every network call needs your consent, and everything the agents do goes in a tamper-evident audit log.

Everything is analysed and stored on the device and sealed with AES-256 under a key kept by the phone's secure hardware. There is no account, no ads, no tracking and no server holding your health data. It works in English and Hindi, with UK, US and India health packs. In the app, More → Try each feature plays guided demos with a sample family.

AuraHealth gives wellness information only. It is not a medical device, and camera, voice and motion results are estimates.

FaunaVox listens to animal sounds and tells you what they likely mean: your dog's whine, your cat's yowl, the bird in the garden or the deer barking in the forest. Recognition runs on the phone or in the browser, so sounds are never uploaded.

It has two modes. Pet Mode helps owners understand their own pets: it learns each pet's voice, keeps a mood diary and a health log for the vet, and can watch a pet left home alone. Wild Mode names animals by their calls, warns about dangerous animals and reads the forest's alarm calls, with stations and alert groups for people who live or work near wildlife.

Every result says how sure FaunaVox is and why. Meanings come from animal-behaviour research and are shown as likely moods, never word-for-word translations, and the accuracy for every animal is published. There is no account, no ads and no tracking. It works in 9 languages: English, Hindi, Marathi, Bengali, Tamil, Telugu, Kannada, Malayalam and Spanish.

FinancialScan is a fraud prevention and detection platform for Indian banks, built with co-operative, regional rural and small finance banks in mind. It is made of separate features: each bank subscribes only to the ones it wants, and features share data only when the bank allows it. Every threshold, rule and piece of wording can be changed from the console, with four-eyes approval.

At account opening it checks the applicant's identity (Aadhaar QR and Aadhaar-app credentials, PAN), looks for deepfaked selfies, videos and voices, risky phones and SIMs, and the signs of a mule account, then gives a GREEN, AMBER, RED or BLOCK decision with its reasons. Risky applications go to a reviewer from another branch, who gets a numbered case file and drafts of the customer notice, police report and STR. A guided interview in the applicant's language covers branch, business-correspondent and online openings.

After opening, it watches the first 90 days for accounts that are rented or sold, checks a change of mobile, SIM or nominee before it is applied, warns customers at the moment they press Pay, manages debit holds under RBI's procedure, and runs RBI's compensation framework for small digital frauds (from 1 January 2027), including signed claims between banks. Small banks can share mule identifiers with each other privately. It also defends against bots and AI agents, and lets a customer give an AI agent strictly limited permission to pay (Google's AP2 and Mastercard's Verifiable Intent).

Every decision gets a receipt signed with a classical and a post-quantum key, which police or auditors can check without an account at the bank. The deepfake models run on the bank's own servers, and nothing goes to a hosted AI unless the bank switches it on. A privacy console covers India's DPDP Act. It is tested by 65 scripted personas (306 of 306 checks passed) and a demo bank that walks through the whole flow with sample applicants.

FlowIQ turns a small business's existing CCTV cameras into numbers the owner can act on: how many people came in, how long they waited, who was served and who walked out, the busiest hours, and how the staff's time is spent. It runs on a small computer on the premises, works with no internet connection, and video stays on site by default.

Nothing has to be set up for each venue. It learns each venue's own normal from how the room is used: which people are staff and which are customers, where the waiting area and service points are, when the venue is open, and what an unusual queue or wait looks like there. Zones can also be drawn on the camera picture: a waiting area, the service chairs, or a mirror, so reflections are not counted as people.

Every number says what it rests on, and when the camera cannot see something, FlowIQ says so instead of showing a zero. A salon whose door is out of shot is told its waits cannot be measured, not that nobody waited. Safety alerts for fire and smoke, falls, guns and knives come with a clip and a signed, tamper-evident evidence record, an incident report for an insurer, and Content Credentials on the clip.

It can also run a home or small building: Home Assistant and Zigbee devices, rooms that combine cameras with presence sensors, and rules such as "turn the heater off when the room has been empty for 15 minutes", with electricity prices from a flat rate, Octopus Agile or Home Assistant. A voice or text command can never unlock a door, open a cover or disarm the alarm.

It has been tested on live public cameras in salons, barbershops, pizza kitchens, a car workshop and a hotel square. A small two-core server handles about two cameras. Theft detection is not claimed. It is a private project and is not for sale.

HaggleHub turns shopping around. A buyer posts what they want, from a laptop or a pair of trainers to a watch, an appliance or a trip, with the most they will pay and how long the request stays open. Verified local shops then compete for the sale.

Shops never see the buyer's budget unless the buyer chooses to share it. They see the lowest bid so far and can lower their price to beat it, so the price keeps falling until no shop can go lower. The buyer can counter-offer on any bid, and the shop accepts or comes back with its best price. Shops only see requests in the categories they are approved for, and shops rated under 4.0★ are blocked until reviewed.

Next to the shop bids it shows real online prices: eBay, Best Buy and Sovrn Commerce for products, and Duffel and Aviasales for flights, once their API keys are added. Without a key it shows plain search links and never invents a price. A watcher re-checks the prices every 15 minutes while the page is open and alerts when one is at or under the budget. The buyer pays the chosen shop directly; HaggleHub never holds the money.

This is a working prototype: requests and bids are stored in the buyer's own browser, the competing shops are bots that price from a catalog of store prices, and shop verification is simulated. The web demo has no price keys connected yet, so it shows search links instead of online prices. It is tested by 10 buyer and 5 shop personas run against the real auction code (71 of 71 checks passed) and 21 checks of the price sources.

JanHelp helps people in India with money and legal problems. Its agent part finds the right loan, credit card or insurance for a person and helps when something goes wrong with one: a refused claim, a wrong charge, a mis-sold policy. Its do-it-yourself part has legal, tax and government tools and calculators. It drafts the notice or complaint, tracks the deadlines it starts, and gives the steps to file it on the government's own portal.

Nothing is made up. Rates, fees, eligibility rules and premiums are read from each lender's, card issuer's and insurer's own published pages and checked again every night, and every answer says where it came from. No AI chatbot writes the answers. AI assistants such as Claude can use the same tools through JanHelp's MCP server.

It works in English, Hindi, Marathi, Tamil, Bengali and Telugu, and needs no account. What a person gives it is encrypted wherever it is kept: on the server, in the browser, and end to end in an optional passkey account. Each of its 110 features says who it is for and has a step-by-step guide, and 90 of them can be tried with a worked example.

Applying through JanHelp is in test mode for now: it records the application and prepares the papers, but nothing is sent to a provider yet. It gives tools and facts, not legal or financial advice.

KidPoints is where a child learns and practises, from Reception to Year 13, following the curriculum and exams of the UK, the US or India. A child practises skills, reads a lesson for each one, works through courses and sits mock exams. They earn points for studying, within the daily and monthly limits of their parent's plan.

Questions come from about 5,600 skills. Each skill has its own question templates, banded easy, medium, hard and expert. Questions are generated in advance, checked by automated gates and stored before any child sees them. There are over 5,100 lessons and about 400 courses: school years by subject, the 11+, GCSE, A level, SQA, SAT, AP, CBSE, JEE, CUET and more. Each course is mapped to its official syllabus, with mock papers sized to the real ones and mark schemes that show every method and accuracy mark.

Learning also happens through play. In about two dozen open-source 3D games the questions are part of the game: the answer is a ring on a race track, a numbered pool ball or a creature to blast. There are walkable 3D Knowledge Towns built from the child's school year, with a teacher in every hall, and Hero City, where a child trains twenty superpowers. Branching stories and chess against a clock carry questions too. Working labs cover physics, music, Logo and Python coding, geometry, circuits, the periodic table and the solar system. In Debate Club a child argues a motion against a tutor.

Parents see progress and monthly mastery reports. Teachers teach from any course or their own edited copy, set homework, worksheets and tests, mark mock papers and plan revision for a class. A teacher can also add adults as students.

By default no child's words are sent to an AI model, and Debate Club's tutor is written in advance and stores nothing a child writes. The Android app is the same app as the website. No real children use it yet: it is in testing before launch.

Kinfolk is one emergency fund for a whole family spread across the US, the UK and India: children, adults and elders. Family members add money in their own country, and anyone in trouble, anywhere, can get local money from their phone with nobody else awake. Bigger amounts go to a family vote.

In the US and UK the money is held as USDC in a family Safe that only the family's passkeys control. Licensed providers such as Coinbase, Robinhood, MoonPay, Transak, Ramp and Stripe take top-ups and pay members to their own account, card or bank, so Kinfolk never holds the money. In India the money stays in rupees, in a bank account two elders keep, used through UPI Circle. For now the India side is manual: payments happen in the UPI app and are recorded with their 12-digit reference.

The server enforces the rules, and every action is signed with a key on the person's own device. Up to $60 a day needs no one's say; up to $2,500 is paid instantly in an emergency and the whole family is alerted; above that, a request needs one or two approvals and is paid automatically once approved. Every top-up is kept under the giver's yearly gift-tax line in the US, the UK and India.

This is a working prototype: money movements, providers and notifications are simulated. A guided tour in the app shows every feature and can do each step for you.

Nexus Travel is a business travel app for companies, built India-first. A traveller says where they need to be, typed or spoken, in English or Hindi. Nexus drafts the flights, hotel and transfers around the meeting, checks them against the company's travel rules, and sends the trip for approval only when the rules say so.

It works with however a company books today. Companies on Concur keep it, and Nexus adds pre-trip checks, money back and GST credit. Companies that book through local travel agents send each request to their agents by email or WhatsApp: Nexus reads the quotes, compares them with each other and with the online price, and shows when booking yourself would be cheaper. Companies that book on set websites such as MakeMyTrip get those sites opened already filled in, and the confirmation email becomes the booking.

Every trip shows its stage, who has it now and what happens next. Travellers can change or cancel a booked trip and see the cost before anything changes. Finance sees spend, money recovered and GST credit, and admins switch features on or off for their company.

Data is encrypted on the way and where it is stored, each company's data is kept apart, and roles decide who sees what. The Android app opens the same service and has a second door to Travel Haggle, for personal trips where agents compete to quote. The demo uses fictional companies: don't enter real personal details.

OmniQA is an autonomous testing platform. Point it at a web app, including single-page apps and Pega, or at an Android or iPhone app. It explores the app, tests it and writes tester-grade reports: what passed, what failed with screenshots and traces as evidence, and a ready-to-run Playwright test suite.

Every kind of testing is opt-in. For each app you choose from 32 kinds of testing, such as journeys, accessibility, security, performance and APIs, or a preset such as smoke, accessibility or full. Nothing runs that was not chosen.

The Test plan page gets what the testing needs from each app's team. Choose the testing and it creates one questionnaire file with only the questions those tests need, each with its answer format and an example. The team fills it in, in any browser and even offline, and saves it. You upload it, check what is missing, and start testing. When the run finishes, everything downloads as one zip.

Passwords never go in the questionnaire. They are typed in at upload and kept in memory for that run only. Integrations are tested with sandbox keys only, and live keys are refused. Security checks are defensive and are not a penetration test, which stays human-led. Load tests stop on their own if the app starts failing.

PetWithBrain is a 3D pet, a dog or one of three original creatures, run by a brain modelled on how real animals learn. Nothing is scripted: every trick, word and habit comes from what happens to it. A reward it didn't expect teaches it the most, sleep makes the day's learning last, its feelings speed up or slow down its learning, and each pet is born with its own temperament.

You teach it with treats, praise, touch and your voice. It learns your routine, the people it meets, the names of its toys and the words on its talking buttons. Mind → How it learns shows what is helping or holding back its learning right now and what makes its brain different; the learning journal shows everything it is learning, day by day.

Its brain runs on the phone and works with no network. Backups are end-to-end encrypted with a recovery code only you have, the service is HTTPS only, and Android's own backup never takes your sign-in or recovery code to the cloud.

For testers, You → Try every feature explains each feature and shows where it is, with a demo pet that already has three weeks of history.

QuantaProof examines an uploaded image, audio clip, document or video with a panel of independent forensic tests: metadata and file structure, compression history, content credentials (C2PA), a tamper localiser that marks edited regions, and learned detectors for AI-made pictures and cloned voices. It reports what each test found rather than a bare verdict.

Each test's false-alarm limit is set from real genuine files of the same kind (camera photos, WhatsApp forwards, voice notes, phone recordings, scans and identity documents), and the whole panel shares one limit, so about 5% of genuine files are flagged by design.

Measured on files it was never trained or calibrated on: AI voice clones in Indian languages, across four channels (as published, voice note, phone call, wideband call), 317 of 400 caught (79%) while 17 of 480 genuine FLEURS speech files were flagged (3.5%); AI images from 10 generators it never trained on, 274 of 400 caught in the standard scan (68.5%) and 340 of 400 in the advanced scan (85%). Video detection is still weak, and false alarms on genuine Indian-language speech are being re-measured.

A research prototype: several of its detectors carry non-commercial licences.

TerraGuard is a family safety app built for India that gets an SOS through even when there is no mobile data. Hold one button, or press the power button three times with the phone in your pocket, and your family, your emergency contacts and, if you switch it on, 112 get your exact position, battery and what's wrong. With no data it goes by text message, by satellite on phones that support it, or phone-to-phone over Bluetooth. A family member with a basic phone can raise the alarm with a missed call or a "HELP" text.

It also acts when you can't ask for help. Fall and car-crash detection run on an on-device AI model, "walk me home" and ride check follow you until you arrive, and a phone snatched from your hand sends its location to your family. Family live location, place alerts and check-ins are end-to-end encrypted, and sharing is always visible and can be paused.

Everything the app saves is encrypted on the phone with a key held in its secure hardware, the server can't read family locations, check-ins or messages, and there are no ads. A practice mode lets anyone try SOS, countdowns and alerts without reaching anyone.

TradeFlow AI is invoice discounting for Indian supply chains. A supplier's GST e-invoice is matched line by line against the buyer's purchase order and goods receipt, pulled from SAP, Tally or GeM. Before any money moves it is checked against India's own systems: the IRN on the e-invoice portal, the e-way bill, FASTag toll crossings on the route, the buyer's decision in GST's Invoice Management System, CERSAI, and a shared registry that stops the same invoice being financed twice.

The buyer approves verified invoices in one signed step, and banks and NBFCs bid in a reverse auction to pay the supplier early. The amount funded is net of TDS and of the GST TDS government buyers deduct, with GST held back until the buyer can claim it. Small buyers that no bank has a limit on are priced by credit grade, and CGTMSE or NCGTC guarantee cover lowers the bid. Suppliers can draw money on WhatsApp in Hindi, Gujarati or Tamil, by text or voice note, after a Key Fact Statement and an OTP.

Buyers see the 43B(h) tax at risk and the MSMED interest building up on late payments to small suppliers, and get MSME Form I. Public sector companies get their GeM orders and their duty to settle MSME invoices through TReDS handled. Financiers get supplier KYC and land records through RBI's Unified Lending Interface, credit memos built from Account Aggregator and GST data, e-NACH collections and guarantee claims.

This is a working sandbox: the companies are fictional and the government and bank systems are simulated, so no real money moves. The demo page runs 32 journeys for 10 personas on the server and shows what each step returned; the API is documented at /docs. Running it for real needs a TReDS licence or lender partners, and access to each government system.

The calculator works out a UK Self Assessment tax bill for the tax years 2023 to 2024 up to 2026 to 2027. It covers Income Tax in England, Wales, Northern Ireland and Scotland, National Insurance, Capital Gains Tax, student loan repayments, the High Income Child Benefit Charge and pension charges. It then shows what to pay and when, including payments on account. Every rate and allowance comes from GOV.UK or legislation.gov.uk, and the pages follow the GOV.UK Design System.

It can be used in 3 ways. Customers can answer one question per page and see the bill with the working. Another app can put the question pages inside itself. It sends the customer with what it already knows, and gets them back with the tax and what to pay when. Or software can call the API, for the full calculation or one part at a time from the least it has, such as pay alone.

An administrator chooses which of 22 parts of the calculator are used, for example basic Income Tax only. They also choose which ways are open, where answers are read from HMRC records instead of asked, and how much working the result shows. Each organisation using the service has its own settings, managed by the service administrator or its own administrators. Customer accounts keep a forecast up to date during the year, and an optional feature works out Budget Payment Plan amounts.

It is tested in 5 ways: 121 hand-worked customers, 42 end-to-end journeys, 376 threshold checks, 20,000 random customers checked against a separately written calculation, and 97.8% code coverage. In all, 1,039 automated tests pass. The test site uses dummy versions of HMRC, GOV.UK One Login, Notify, DVLA and a partner app, with 14 test customers and 3 example organisations.