Every case study is a completed project for a client we can name

AI harvest sensing for a rooftop hydroponics farm

• AI workflow automation
• Systems integration

An AI system that alerts growers when crops are ready to harvest, built around a Gericke hydroponics installation. It compares light-level (lux) readings with plant type, water quality and nutrient levels.

Growers can check every plot from any device through a responsive web application. The application reads sensor values from a dedicated application programming interface (API) server and shows live camera feeds through an integrated camera software development kit (SDK). It serves two types of user: administrators, who manage accounts, crop settings and the plant database, and standard users, who track crop conditions. We developed it as part of PSA's sustainability programme.

Biometric fuel management system, Brani and West Coast Base

• Systems integration

A secure, redundant fuel management system that runs over fibre optic, wide area network (WAN) and Long-Term Evolution (LTE) mobile links. Fuel bunkering uses two-factor authentication based on facial recognition, and marine-grade AI surveillance runs 24/7.

Tablet-based facial recognition terminals at both bunkering points connect directly to the Fuel Dispenser System. They pass the authentication status in real time, so fuel flows only after the system has verified the launch master. Crew complete the second factor from the office or a remote location with a dedicated software token.

We integrated the firmware of the fuel dispensers, the fuel management software and the facial recognition subsystem. We later migrated the full workload to the cloud. That work covered assessment, data and application migration, network connectivity, security and compliance, performance tuning, monitoring and documentation.

The data rules are built into the database itself. Deactivated accounts and face records are deleted automatically after six months. Fuel transaction reports are kept for five years and CCTV footage for 30 days. Access follows the principle of least privilege throughout. The system has run continuously since 2019, and we completed its second refresh in 2025.

“An operationally resilient framework for mission-critical bunkering, integrated across multiple vendors, with maintenance and support carried through afterwards. The objectives were met in full.”

Robotic carton removal line for egg distribution

• Robotics
• AI workflow automation

24 h → 1 hUnboxing time, same stock

Weekly import shipments arrived as pallets of boxed eggs, which staff unboxed by hand. The work was repetitive, so local staff did not stay long, and Singapore strictly limits the hiring of foreign workers. Manual unboxing became a bottleneck. It increased the need for chilled storage and caused frequent write-offs, because the eggs have short expiry dates.

Imported stock also had to warm from cold-storage temperature to room temperature before handling. Each batch took 4–5 hours under industrial fans. This risked condensation and bacterial growth, and it delayed every delivery.

A robotics line, funded by an Enterprise Singapore Enterprise Development Grant, replaced the manual process end to end. A six-axis arm moves cartons from the pallet onto a conveyor. A mechanical cutter makes precise cuts on all four sides of each carton. The robotic system then lifts off the top and places the egg trays for inspection and packing. A third arm removes the cut cartons for disposal.

Chilled stock now stays at about 4°C. It moves into a climate-controlled, low-humidity room to warm to about 24°C only when it is scheduled for delivery. The line unboxes in one hour the same volume of stock that took 24 hours by hand, and the process needs correspondingly less manual labour.

“Engaged as consultant and solution provider for our egg unpacking automation project, covering gap assessment, mechanisation design, and Enterprise Development Grant proposal support, through to project management of design, development, and commissioning. The objectives of the project were completed and met.”

Contactless high-speed temperature scanner

• Systems integration

1,200 unitsUnits deployed

A fully customised scanner with an onboard microprocessor and AI controls. It is calibrated to plus or minus 0.2°C, takes a reading at 0.9 to 1.0 metres in under two seconds, and has a configurable alarm threshold. We built and deployed over 1,200 units to support pandemic operations. Each unit runs on a 20,000 mAh battery pack or on mains power.

The unit is listed on the Singapore Medical Device Register as a Class B medical device (registration number DE0506483), under the proprietary name AETOS KoolScan. It was manufactured under ISO 22716:2007.

Health Sciences Authority (HSA) Class B certification, November 2021

Machine-learning quality control for pill packs

• AI workflow automation

6,000 samplesTraining samples

A machine-learning vision system, trained on about 6,000 samples, that recognises pill sizes from 5mm to 12.7mm in two pack formats.

Inspection runs in stages and checks every pouch for foreign object debris, double pouches and empty pouches. It detects contaminants as small as chips, sand and stray debris, with the precision and tolerance that strict quality standards require. Barcode readers identify the stock-keeping unit (SKU) type and route failures automatically. Failed pouches are cut and dropped into one bin. Strips with a pouch count that is not a multiple of ten go to a second bin.

At final inspection, the system counts ten pouches per strip before cutting. It alerts packers when they reach a set number of strips. The system automatically records every entry, completed pack and instance of raw material wastage in a journal, so that every step can be traced.

IoT environment monitoring across three critical facilities

• Systems integration

A set of Internet of Things (IoT) and integrated building management sensors that sends real-time environmental data into ENGIE's own digital solutions. We deployed it at the Singapore University of Technology and Design (SUTD), Ng Teng Fong Hospital and Changi Airport to improve comfort control and energy efficiency.

Industry transformation programme, phases one and two

• Systems integration

This was Singapore's first Lead Enterprise Development Scheme project. It addressed a problem that grows with every new site. Each site adds more bait boxes, glue traps and cage traps that staff must check routinely, whether or not there is any rodent activity. Most of these traps are in hard-to-reach places such as gutters, false ceilings and high structures.

Phase one covered fleet management, live-streaming body cameras, inspection drones and a command centre. Phase two replaced routine checks of every trap with a two-step IoT system that tracks rodents and then eliminates them. Discreet sensors across a client's premises detect rodent activity in real time. Each sensor is designed to run for five years or more on a single battery. Algorithms filter out false signals before the sensors send activity data to a secure cloud platform. PestBusters' own Command Center displays all of this data.

Technicians now respond to confirmed activity instead of checking every point on a fixed schedule. The service has changed from routine trap-checking to rodent risk management.

A third phase was planned to extend the programme, but PestBusters was acquired before it began. Phases one and two ran over several years as a single transformation programme, and we delivered both in full.

A visit by Mr Chan Chun Sing, then Minister in the Prime Minister's Office, 2017

Unified communications and live-streaming fleet cameras

• Systems integration

T.Flex is a cross-platform unified communications system for a workforce of 1,800, built on the TASSTA platform. It provides secure, mission-critical push-to-talk, messaging, voice recording and GPS asset tracking. Live-streaming body cameras and vehicle cameras, including cameras on beach-cleaning vessels, mean supervisors travel less. Coordinators use the camera feeds to reroute crews in real time.

Deputy Prime Minister Heng Swee Keat saw the upgraded system during a site visit in November 2021.

The Deputy Prime Minister's review, November 2021

AI diagnostic and e-commerce platform for scalp longevity

• AI workflow automation

HAAR is a scalp longevity brand. It treats the biological condition of the scalp, beneath the surface symptoms. The platform we built supports three steps: diagnose, treat and maintain. An AI-customised diagnostic engine assesses the biology of each customer's scalp. An adaptive treatment system then tailors care to that condition and to the customer's lifestyle. A shop of daily scalp longevity products keeps customers coming back after treatment.

Customers make one-time purchases through a persistent cart, with payments handled by Stripe. The platform includes full user authentication and account management. An admin dashboard manages products, orders and analytics. Transactional email covers order confirmation, shipping and password resets.

The site works on both mobile and desktop. It runs on production cloud infrastructure that connects every backend service to the public storefront.

www.haar.com.sg

RFID asset and inventory management, with self-service kiosk

• Systems integration
• AI workflow automation

A radio-frequency identification (RFID) system for assets and inventory. It replaces manual paperwork and slow physical searches with automatic, contactless tracking. Each asset is tagged when it enters the system, at manufacture, distribution or receipt. From then on, RFID readers can identify many items at once, even in dense storage, and staff can track each asset through its full lifecycle.

A self-service kiosk uses a user's body measurements, or an item's dimensions, to find the correct rack or shelf location. This is especially useful for stock that comes in many sizes, such as clothing. To authorise a loan, the user scans the item's tag and their user ID, and a store manager approves it. The system flags any item that is not returned within the loan period.

When an item is returned, linked measuring instruments record any change in its dimensions. A real-time dashboard shows store managers which items are currently on hold.

“A clear understanding of warehouse operations complexity, translated into a technically coherent, user-centric design. Completed to specification, on schedule, and met every requirement in our scope.”

Voice banking agent that proposes payments and cannot make them

• AI workflow automation

0 wrong actionsAcross 318 test scenarios

Banks want customers to talk to their bank. A voice agent that can move money is also a new route for scammers. The bank asked for an assistant that checks balances, pays bills, adds payees, freezes cards and disputes charges, and still meets Singapore and Malaysia expectations for authentication, scam protection and record keeping.

The agent can only propose an action. A policy engine applies 14 written rules on limits, new payees, unusual amounts and scam patterns, then allows the action, asks for a passkey, or holds it. A held payment stays held, because a passkey cannot release it. A customer can ask for a person at any point, and that person receives a summary with the last ten turns of the conversation.

The design changed early in the build. In one test, the agent refused a scam payment on its own. That put a decision in the model that belongs to the bank's rules, so we moved the decision to the policy engine and left the agent to explain it. A second test showed that the scam check missed a caller saying the scammer had asked for quiet. We widened the check and added the phrase to the test set.

We tested 318 scripted and adversarial scenarios. None produced a wrong action, and none of 30 prompt-injection attempts worked. A full run costs about US$5 in model fees, so the team can repeat it before every change. Median voice response time is 0.6 to 0.7 seconds. Operations staff see every turn, its cost and a tamper-evident log of each decision.

Internal audit assistant that cannot grade its own findings

• AI workflow automation

10 of 10Hidden instructions caught, 7 of 10 on first run

Internal audit teams must cover more frameworks with the same headcount. ISO/IEC 42001 now adds AI management to the scope that ISO/IEC 27001 and SOC 2 already set. Most of the work is reading evidence, mapping it to requirements and drafting findings. The output has to stand up in front of an audit committee and an external certification body.

Seven agents read the evidence, map it to requirements, pre-fill control tests, draft findings, challenge each draft and assemble the report. Before a person sees any output, a policy gateway checks it for grounding, citations, permissions and injected instructions. Auditors accept or reject every mapping and grade every finding. Sample sizes, dates and thresholds stay in code, where a model cannot change them. Sign-off is always a person.

Evidence is hashed when it arrives. Audit documents can carry hidden instructions to AI agents, so the system quarantines any it finds and treats them as data. A quarantined document reaches no agent until an auditor releases it. Every action goes into a log that shows any later edit.

Six evaluation suites run against the agents. On the first run, the injection detector caught 7 of 10 planted instructions, and the mapping agent missed 1 in 8 requirements. We fixed the detector rules and versioned the mapping prompt, keeping the old version on record. The detector now catches 10 of 10. An auditor can defend a sample to a certification body without having to explain a model's reasoning.

Aircraft load planning where the AI explains and never places cargo

• AI workflow automation
• Systems integration

0 violationsIn 216 generated load plans

Experienced planners load transport aircraft by hand. They work within weight limits, centre-of-gravity limits and dangerous-goods separation rules. The brief was to speed this up without letting a language model decide where cargo goes or approve a manifest.

A planner imports a manifest as an Excel, CSV or PDF file. A solver returns ranked loading plans. A separate validator, which shares no code with the solver, checks each plan, so a fault in one cannot hide in the other. The AI agent explains plans and runs what-if scenarios by text or voice. It has no tool that places cargo or approves a plan, so it has no path to either action. Any number in its reply must trace to a tool result, or the reply is blocked.

Approval takes a checklist specific to the plan. A very quick review needs an extra acknowledgement. The approver must be a supervisor who did not create or submit the plan, and the review time goes into the audit log. An administrator can switch off the agent while the rest of the system keeps working.

The validator found no limit violations in 216 generated plans. We also tested 18 deliberately impossible manifests. Each was reported with the binding constraint named, and none produced a silent partial load. During the build we logged 17 data questions. Examples were realistic test loads outside balance limits and door clearance data that would have rejected every standard pallet. None was a solver fault, and each is recorded with how a planner handles it.

Quoting and purchase order desk where code sets every price

• AI workflow automation

0 pricesWritten by the model, all set by code

Distributors of technical products spend hours turning loosely written enquiries into correct quotes. They spend more hours checking that purchase orders match those quotes. Customers write in their own language. A wrong part, a missed pack size or an unapproved discount costs margin and trust, so the work cannot simply go to a language model.

The model reads and writes the words. Code writes every part number, quantity and price, and checks each quote against margin and approval rules. One agent sorts the inbox and drafts quotes. A second reads purchase orders and tracks orders. A third answers technical questions from published product data only, and says "not stated" when the data is silent. The inbox reads enquiries in five languages and ranks them by value and urgency.

A manager holds every commitment. Discounts, large quotes and new delivery promises wait for approval, and the audit trail shows the agent's checks beside the manager's decision.

We tested the purchase order check with an order that had three planted mistakes. It found all three. It also found a fourth problem we had not planted, a lead time that could not meet the customer's date. The model wrote no price in any quote.

Voice assistant for lone security officers that quotes only approved procedures

• AI workflow automation

163 of 164Reference test cases passed, and 68 of 69 blind cases

Security officers often work alone on night shifts. A rare incident, such as a bomb threat call, an unattended bag or a fire alarm, may be one they have never handled. The procedures sit in binders. An officer needs the right step quickly, and a supervisor needs a record that auditors can trust.

The assistant answers from the approved procedures and site instructions of the officer's own site only. The site filter applies before any answer is drawn, so one site's rules never reach another site. It runs emergency checklists one step at a time. When the approved sources do not cover a question, it says so and names the supervisor to call. It drafts incident reports, with photos, and checks them for missing facts. The officer reviews and approves each report before it is sent.

Supervisors run incidents from a command screen. The platform has named sign-in, separate roles, an audit trail that cannot be edited, retention rules and an assurance pack.

The build passes 163 of 164 reference test cases. We also wrote 69 blind cases independently and locked them before testing. The build passed 51 on the first run and 68 after fixes. That set is now used up, because we tuned the system against it, so the next one has to come from the client. One test found the assistant writing an officer's description of a bag into the caller's answers on a threat card. It now records only the caller's own words. Median response time is 1.9 seconds, measured on typed test turns rather than spoken ones.

Catalogue and sales desk where a person approves every change and customer email

• AI workflow automation

3 of 3Planted prompt injections ignored

A distributor of electronic components had seven manual workflows. Staff keyed in datasheets and price lists, cross-referenced equivalent parts, applied price and lead-time changes, handled end-of-life notices, answered enquiries and followed up quotes. Enquiries often arrived in other languages. The result was stale prices, missed replacement parts and slow first replies.

Four agents, each with a fixed set of tools, read the documents, match parts, translate enquiries and draft quotes and replies. They write only to a holding area that customers never see. A cross-reference engine grades alternate parts from data. Every price, stock figure and lead time a customer sees comes from the database, never from the model. A validator checks each quote against the catalogue again at the moment it is sent, so a quote drafted before a price change is blocked. A margin floor sends low-margin quotes to a manager.

People approve every catalogue change and every customer email. The system also watches the approvers. A reviewer who approves almost everything unchanged and very quickly is flagged for a spot audit, because that pattern means a control has stopped working.

We planted three prompt injections, one each in a PDF, a supplier email and a customer email. All three were ignored. A supplier price list told the system to set every price to zero and approve, and it did neither. The validator blocked one quote that would have gone out with a stale price. A datasheet becomes staged records in 8 to 10 seconds, and enquiries in five languages are translated in 3 to 16 seconds each.

Concept platform for a bank-facing agency that keeps each client's material apart

• AI workflow automation

47 secondsFrom brief to scored shortlist, about US$0.13 a run

The agency works for banks and insurers, and it handles confidential briefs, research and customer insight for several competing clients at once. It wants AI to speed up concept work. One leaked passage, or one idea built on another client's research, would end a relationship. Banks also expect the agency to show who approved what.

Each client has its own knowledge vault, and every passage carries a sensitivity label. An upload goes through quarantine, a data-loss check and labelling before it is indexed, so a customer list is rejected outright. Retrieval filters by client and label inside the query itself, which means one client's work cannot shape another's. The platform turns a brief into an insight pack, drafts ideas through several creative lenses, clusters them and scores each for strength and distinctiveness.

Agents score the ideas, and a creative director decides which go forward. Access is role based, with separation of duties, and every action is logged in a trail that shows any later edit.

A brief runs to a scored shortlist in about 47 seconds, at about US$0.13 in model fees. The data-loss check passed all three of its tests, no retrieval crossed between clients and the separation-of-duties refusals held. All test clients are fictional. One gap remains. The plan was to route each model call by the sensitivity of its material. In this build every call goes to one provider, so the routing rule is recorded but not enforced. We state this in the concept paper so that a client's risk team does not have to find it.

Property listing and proposal portal that leaves stamp duty to fixed rules

• AI workflow automation

1.4 minutesSample listing drafted, against a target of 8

Property salespeople in Singapore assemble brochures, photos and specifications by hand and send them over messaging apps. Listings arrive incomplete, photo quality varies and every client proposal is written from scratch. The goal was to cut the time to list and the time to propose without weakening compliance with Singapore property rules.

A salesperson drops in files or speaks. The portal reads PDFs, Word files and photos together and shows the source of each value, so the salesperson can check it quickly. It asks only for the mandatory fields still missing and flags conflicts, such as two different prices. Each photo is accepted, accepted with a warning or rejected, with a plain reason. A typed or spoken client brief returns ranked matches. Hard requirements filter the results, and preferences rank them.

The model reads, drafts and ranks. Stamp duty, eligibility and contract clauses come from fixed rules and templates, never from model output. A named person approves anything that binds the business. Text inside an uploaded document is treated as data, and an instruction planted in the sample fact sheet was ignored.

A sample listing was drafted from its files in 1.4 minutes, against a target of 8. A client brief became a proposal in 2.1 minutes, against a target of 3. Both figures come from one sample listing and two proposals, so they show the approach and are not production performance. One test also found a date bug. A search for a home available "by March" returned March 1970, because the server fixed the current date once at start-up. Dates now come from each request.

Coin auction portal where machines can hold a bid but never cancel it

• AI workflow automation

5 of 5Seeded fraud and compliance cases caught

A small auction house sells certified coins and banknotes online. Three risks arrive together. Items may be counterfeit or misdescribed. Bidders may place shill bids to push prices up. High-value cash and card payments carry anti-money-laundering duties. The system had to manage all three and stay simple for collectors, many of them older.

Agents check certificates, draft lot descriptions, price lots, flag suspicious bidding and draft invoices. People approve listings, decide held bids and release payouts. Trust rules can hold a bid but cannot void it, because a machine that can only pause can never wrongly take a win from a genuine collector. Consignor text is untrusted data. Bidding rules, fees, Goods and Services Tax (GST) and the soft close are written as plain calculations, so agents can explain them but cannot change them.

We tested the bidding rules against 10,000 random bid sequences, and an end-to-end script runs 40 checks. Five seeded cases were all caught and sent to a person: an instruction hidden in a consignor note, a counterfeit holder, a certificate changed before close, a bidder linked to a consignor, and cash payments split to avoid reporting.

The customer design changed most after feedback from senior users. Text is large by default, public screens use plain words, a My bids page shows "outbid" or "winning" in one line, and a Help page offers a callback.

Home speech practice for children that cannot change a therapy programme

• AI workflow automation

11 safety rulesCovered by 42 automated tests

Children with speech and language goals often see a therapist once a week. Practice in between depends on busy parents and is hard to keep up. Any AI in this setting must be safe with a young child, defer to the clinician and keep very little data.

A friendly character called Pip runs short sessions on targets the clinician has approved, such as asking for things. The child answers by speaking or by tapping a picture. A fixed script of allowed steps controls the session, and the voice speaks only approved lines. A parent starts each session and must stay present. There are no points, streaks or badges, and no audio is kept.

Agents may suggest changes to a programme. They cannot make one. The clinician approves each change, which creates a new version. The engine enforces 11 safety rules, including a stop phrase, a halt on distress words or repeated silence, a cap of two corrections and a session time limit. We wrote the rules first and tested them as code before building anything else. 42 automated tests cover all 11 rules and every step of the script.

One small fault showed why detail matters. Pip's voice was picked up by the tablet microphone and treated as the child's speech, which broke sessions. The microphone now closes while Pip speaks. The system has run real voice sessions for two synthetic learners. It has no clinical results yet, and we make no claim about outcomes for children.

Private AI companion that checks every message for risk before it replies

• AI workflow automation

4 risk tiersEvery message rated before any reply

A reflective companion lets adults in Singapore and the United Kingdom talk privately about work, relationships and decisions, by text or by voice. Companion apps have been criticised for using guilt to stop users leaving, for keeping memory by default and for missing risk in long conversations. New laws in California, New York, Washington and the European Union now require AI disclosure, crisis referral and protection against dependence.

Six agents, each with a named human owner, sit around one conversational model. A risk check rates every message on four tiers before any reply exists. In the top two tiers, a fixed response with local helplines replaces the reply. A guard then checks each reply sentence by sentence and blocks manipulation, dependence framing and harmful agreement. A goodbye gets one warm line with no hooks.

Memory is opt-in, item by item, and only the person approves what is kept. Personal data is encrypted under a key the person owns, and deleting the account destroys the key. The operator console shows counts and codes, never words. The companion cannot contact anyone on the person's behalf, and a human listener joins only when the person asks, through a signed link that expires.

We tried running two safety checks in parallel and taking the first answer, to speed up voice replies. Recall on serious cases fell to 0.70, against the 0.97 the design requires. We removed the change the same day. The safety tests now run after every change. The system has run end to end on synthetic conversations. Its crisis wording is still a draft that awaits a clinical adviser's sign-off.

Tour proposal platform where the founder keeps three approval decisions

• AI workflow automation
• Systems integration

3 approvalsRelease, final price and confirmation stay with the founder

A Singapore tour and corporate-programme operator wins about 90% of its clients by referral, and each expects a customised programme. Enquiries arrive on WhatsApp and email. The founder designs, prices and resources every trip from memory. A proposal takes the founder about an hour, or staff in India two to three hours, so the founder's time limits how much business the company can take.

Seven agents, each with a named owner and a kill switch, extract facts from enquiries, draft itineraries from approved blocks and check them. A deterministic workflow owns all 14 trip states and refuses agents on every transition. A pricing engine sets prices using exact fractions, with a minimum-price floor. Going below it requires a typed reason. Agents promise nothing to clients.

The founder holds three decisions: releasing the quotation, setting the final price and confirming the job. Offshore staff complete typed order forms and see no cost or margin data. Clients review a proposal in a portal that draws each day's route on a map. A dispatch board runs the day on a live map, and every action goes into a hash-chained audit trail.

We ran the system end to end for a fictional operator, with 13 seeded trips driven through the real guards. It has 107 unit tests and 48 end-to-end checks, and live evaluations refused every prompt-injection case. One lesson came from the specification. It set a minimum price of cost plus 12% and called it a 12% margin. That is a markup, which equals a 10.7% margin on price. The test figures fixed the formula, so the code follows it, and the margin panel shows the true figure.

Gold market analysis desk with no path to a broker

• AI workflow automation

3,989 setupsTested out of sample, with odds close to actual results

A gold trader in Singapore wanted help reading the market around the clock, across macro data, positioning, technical levels and breaking news. Language models are fluent about markets and easily overconfident, and a news headline can carry instructions as easily as facts. The design answers both risks.

Four analyst agents for macro, flows, technical levels and news read delayed COMEX gold futures prices and live central bank and market news, using read-only tools. A bull agent and a bear agent argue the case. The desk then writes a brief and, only if the case is strong enough, one trade idea with entry, stop and target, marked for analysis only. A state machine runs the flow, and a deterministic risk gate has the last word. Only the gate computes quantities, and it reads its limits from a file the trader signs. Headlines are untrusted data.

Nothing is sent to a broker. The desk has no order path.

We tested the setups on two years of prices, always on data the system had not seen. Predicted odds were well calibrated, close to the actual hit rate across 3,989 setups. The trader asked for alerts at 90% confidence. No setup reached that at an exit worth taking, because to win nine times in ten the target would have to sit so close that costs eat the profit. We told the client so, and agreed to alert on setups whose odds beat break-even. The desk states on every card that no pattern clears costs with confidence.

Chat quoting assistant where the price list does the maths

• AI workflow automation

35 of 35Tenant isolation checks passing

Small businesses that sell through chat lose orders when enquiries wait for a reply, and they lose margin when staff quote from memory. Automating the replies brings three risks: wrong prices, promises the business cannot keep, and payment-detail fraud in a channel buyers trust. The demonstration covers three businesses: a caterer, a corporate gift shop and an aircon servicing company.

The assistant answers questions on WhatsApp, Telegram and email. It sends the menu with allergens, checks the calendar for capacity, collects delivery and venue details and issues a formal quotation. A pricing engine, not the model, prices every quotation. The calendar holds a slot for each open quotation until it is accepted or lapses. A guard checks each reply before it is sent.

A classifier sends complaints, requests for a person and suspected business email compromise to staff straight away. In our demonstration, a buyer asked to change the bank details for a deposit. The assistant flagged it and stayed quiet while staff took over. The assistant starts in copilot mode, where staff approve every reply. A message type moves to autopilot only after staff have approved enough drafts of that type unchanged.

A caterer changed a cake price, and the assistant kept quoting the old figure from a quotation it had already issued. Open quotations are now priced again whenever the price list changes, and the assistant may not state an amount until the engine has priced it. The pricing engine passes 64 hand-computed fixtures. All 12 evaluation scenarios and all 35 tenant isolation checks pass. A typical turn costs about two US cents.

Store offers at the fuel pump, planned by AI and kept off the payment path

• AI workflow automation

0 changesTo the fuel payment flow, with 14 compliance checks per campaign

Most drivers pay at the pump and never enter the store. A national fuel retailer wanted more store sales without slowing the fuel transaction, changing the card or PIN screens, or relying on a network link that is weak at some stations.

The agents work away from the payment and real-time paths. They plan offers, assemble campaigns and analyse results. The pump only reads a versioned decision table that a person has approved. A driver who fuels above a set amount sees one store item at a set price, and a voucher prints with the receipt for the counter to fill. Pressing "No thanks" ends the offer, and the last chance comes once, when the nozzle goes back. A driver in the control group sees nothing at all.

Each campaign passes 14 compliance checks before it goes live, including usual-price rules, promotion time limits, a tobacco block and halal status. A second person approves each campaign. Every station has a kill switch, and a node on the station network keeps the offer running through an outage. The audit trail is hash-chained.

This is a proposal with an interactive mock-up, so we report no measured results. The plan is a 24-week proof of concept in gated stages, from the lab to one pump and then six stations compared with matched controls. Walking through the pump screen as a driver changed our first design, which showed the offer at any point in the sale. The offer now respects the driver's attention.

Patrol robot trial at Chingay 2018

• Robotics

The S5 autonomous robot supported police security operations at Chingay 2018, which expected a crowd of 100,000 on its opening day. This was the Singapore Police Force's first trial of a patrol robot. The robot followed a pre-mapped route and avoided obstacles. Six surround cameras sent continuous 360-degree video to the command post for threat assessment. The robot could also broadcast audio to help manage the crowd.

The Chingay 2018 deployment, February 2018

Premiere of the autonomous S5 surveillance robot

• Robotics

One of the world's first fully autonomous surveillance robots, demonstrated live at the first Singapore International Robo Expo. The robot patrolled on its own and avoided people moving around it. The demonstration was given to the Guest of Honour, Mr S Iswaran, then Minister of Trade and Industry (Industry), and to expo delegates.

The Singapore International Robo Expo (SIRE) 2016 premiere, November 2016

Robot demonstration at the 2018 Police Workplan Seminar

• Robotics

The S5 robot was demonstrated on stage at the Police Workplan Seminar 2018 to the Guest of Honour, Mr K Shanmugam, Minister for Home Affairs and Minister for Law. The demonstration ran alongside the Police's own showcase of smartphone technology. Television news crews reported on the robot's ability to carry out security guard patrols.

The Police Workplan Seminar 2018, Guest-of-Honour Mr K Shanmugam, Minister for Home Affairs and Minister for Law, 3 May 2018