BluMargins

The challenge

Property in the valley moves through phone calls, messaging groups and personal networks. Listings go stale, enquiries fall between agents, land is measured in marla and kanal, and buyers outside Kashmir have little to judge a property by before they travel. The firm needed the infrastructure of a national portal without losing the agent-led service it is known for.

The approach

The platform is organised around the people who use it. Buyers, tenants and owners each have their own entry point. Every listing carries its verification status, price per square foot and area in local units. Enquiries, visit requests and listing reports land in seller dashboards and the CRM — not in scattered inboxes.

System

1. 01DiscoverySearch by intent — buy, rent, investment, commercial, plots — with budget, bedroom, furnishing, facing and amenity filters.
2. 02Listing integrityVerified badges, ₹/sq.ft, marla · kanal · acre units, shortlists and side-by-side comparison of up to four homes.
3. 03Owner & builder toolsFree property posting, a seller dashboard, listing management and a leads-and-enquiries inbox.
4. 04CRM portalA separate administrative application for agents and staff, with its own sign-in and agent applications.
5. 05Decision toolsEMI calculator, area converter, price trends, investment hotspots and a builders directory.
6. 06MobileiOS and Android apps with price-drop alerts, saved searches and owner chat — announced on the site as coming soon.

The challenge

Civil-engineering knowledge is scattered across drawings, site diaries, sensor logs, codes and surveys. Design, construction, monitoring and hazard intelligence live in separate tools owned by separate teams, so risks that are visible in the data are rarely seen in time.

The approach

Organise the whole lifecycle around a single data model, with capability delivered as modules on top. Every family draws on a shared intelligence layer — a construction-domain language model and a sensor operating system for purpose-built hardware — and ships through a five-tier architecture from field perception to application.

Delivered in phases

1. 01Foundation
2. 02Design Suite
3. 03Site Eyes
4. 04Custom Sensors
5. 05Disaster Intelligence
6. 06Infrastructure
7. 07Full Platform

Seven phases over 48 months, as published. Live today on a free tier: ConstructOS and RiskViz.

The challenge

Construction and investment decisions are made on sites whose risks are documented — in elevation models, seismic hazard maps, flood records and satellite archives — but never in one place. Gathering them costs weeks of specialist time, so feasibility is often judged on a site visit and a hunch.

The approach

Reduce the question to layers a decision-maker can read. Each hazard is scored independently from real geospatial processing, aggregated into a composite with a stated confidence, and narrated by a local language model that describes only the assessment's own data. The system runs fully offline, with no API keys.

System

1. 01Five scored layersTerrain, flood, earthquake, landslide and soil / construction — plus glacial hazard, air quality and infrastructure access where data exists.
2. 02Area assessmentPolygon areas of interest evaluated on a sample grid with worst-case aggregation.
3. 03Change detectionSentinel-2 NDVI before-and-after at the selected site.
4. 04Growth forecastBuilt-up change over time, from historical series where indexed.
5. 05Executive summaryA local LLM narrates the assessment from its own data.
6. 06Plugins & exportsThird-party risk models; PDF, GeoJSON, KML, Excel and Shapefile output.

The challenge

A consultancy that also builds lives on estimates, daily progress reports, labour attendance, material indents and running-account bills. Held in spreadsheets and paper, those records drift apart, and the gap between what was estimated and what is happening on site stays invisible until it is expensive.

The approach

Model the firm's operation end to end, from the first estimate to the last bill, under role-based access with nine role archetypes and an append-only audit trail. Site data is captured offline-first, so work recorded in the field reaches the command centre when connectivity returns.

Module map · as seen in the application

Core

DashboardProjectsCRM

Operations

Site operationsSite engineerApprovalsInspectionsWorkforceAttendance

Supply chain

MaterialsInventory & BOQVendorsEquipment

Commercial

EstimationFinancials

Assurance

Quality & safetyDocuments

The challenge

A single cross-sectional study can run to hundreds of slices. Understanding the true shape of a structure — its volume, margins and relationship to the anatomy around it — means reconstructing it mentally, slice by slice. That is slow, and it is hard to communicate to colleagues and to patients.

The approach

Treat the scan as a volume, not a stack of images. AI-assisted analysis proposes candidate regions for review; volumetric reconstruction renders them as a navigable 3D model alongside the source slices, close to the moment of acquisition. The system is designed as decision support for the clinician — never as a replacement for clinical judgement.

The challenge

Behaviour change depends on timing and context — exactly the data people least want to leave their device. Most assistants solve this by sending everything to the cloud. The brief was the opposite: an assistant as personal as a coach, with none of its knowledge ever leaving the hardware in someone's hand or on their wrist.

The approach

Bring the AI to the data instead of the data to the AI. Language, vision and audio models are compressed to run on-device, alongside a four-part adaptive learning engine and encrypted local memory. The hardware is designed around the same rule: sensing, inference and storage stay on the device, and it keeps working with no connection at all.

System

1. 01Intervention timingLearns when a person is receptive to a nudge — and when to stay silent.
2. 02Response personalisationAdapts tone and style — supportive, direct, playful — to what motivates each person.
3. 03Goal predictionInfers the goals behind behaviour, including unstated ones.
4. 04Context pattern learningModels daily rhythms — places, routines, social contexts.
5. 05On-device model stackLanguage, vision and audio models running locally; encrypted local memory.
6. 06Hardware & wearablesDevice design for always-with-you, offline-first sensing and inference.

The challenge

Exam practice usually reports a score: right or wrong, a percentage, a grade. It rarely tells a student why marks are being lost — whether a wrong answer was a slip, a gap in knowledge, or a misconception they genuinely believe — and so the next hour of revision is aimed at the wrong thing.

The approach

Treat every answer as evidence about a person, not a tally. Each response is linked to the objective it tests, to the student's stated confidence and to the kind of error it reveals. The twin updates, the next question is aimed where the evidence is weakest, and an objective that holds is re-checked two weeks later. Generated questions stay practice-only until a teacher signs them off.

System

1. 01Point-by-point markingAnswers marked against the mark scheme, showing which marking points were earned and which were missed.
2. 02Evidence modelEach answer tied to its syllabus objective, the student's certainty and the type of error.
3. 03Belief & confidence mapEvery answer sorted into secure, a gap, a lucky guess or a misconception.
4. 04Adaptive generationNew questions generated on the weakest objective, from the basics upward.
5. 05Spaced re-checksAn objective that holds is tested again two weeks later.
6. 06Teacher controlCurriculum design, sign-off for generated questions, class heat-maps by objective.