Services
Six ways we put AI to work.
Each capability below answers a different question about your operation, from what will happen next to what the system it runs on should look like. Most engagements combine two or three.
01 / Predict
Forecasting & Decision Intelligence
Demand, price and risk forecasts your planners can defend. Validated the way they will be used, with every number explained.
Most forecasting projects fail at validation, not modelling. A random test split lets the future leak into training and reports an accuracy that never survives contact with production. We score every model the way it will actually run: train to a date, forecast forward, move the date, repeat.
Price history alone cannot see a turn coming. Supply moves on weather, harvests, freight, currency and policy, so we build models around those drivers. That is what lets a forecast anticipate a change instead of confirming one after the fact.
Every figure ships with its reasons. Driver attribution on each forecast shows which factors moved the number, so the person accountable for the decision can judge whether to trust it.
Where it applies
- Demand planning across SKUs, channels and distributors
- Commodity and raw material price forecasting
- Sales, revenue and cash-flow projection
- Credit, churn and supplier risk scoring
- Scenario planning for weather, currency and policy shocks
What you get
- Forecasts with confidence ranges and the drivers behind them
- Scenario tools for testing a decision before committing to it
- Accuracy reported against what actually happened, every cycle
Built with
- Python
- Gradient-boosted ensembles
- Time-series models
- Walk-forward validation
- SHAP
- Climate and market data feeds
02 / Optimize
Supply Chain & Operations Optimization
A forecast says what will happen. Optimization says what to do about it: how much to hold, where to send it and who to schedule.
Planning teams in most operations still resolve trade-offs by hand, in spreadsheets, every week. Hold more stock or risk a stockout. Send one truck full or two half-empty. These are mathematical problems with real constraints, and they have better answers than experience alone can find.
We model the operation as it really runs: shelf lives, vehicle capacities, shift rules, minimum order quantities, lead times that vary by supplier. The optimizer then searches for the plan that meets every constraint at the lowest cost or the highest service level.
The output is a recommendation, not an instruction. Planners see the proposed plan, the cost of each alternative and the constraint that is binding, then accept it, adjust it or override it.
Where it applies
- Inventory and safety-stock optimization
- Route and delivery planning
- Production and workforce scheduling
- Procurement timing and supplier allocation
- Warehouse and distribution network planning
What you get
- Recommended plans your team can accept, adjust or override
- Cost and service trade-offs made explicit
- What-if comparisons before any commitment is made
Built with
- Mathematical optimization
- Mixed-integer programming
- Google OR-Tools
- Simulation
- Python
03 / Automate
Intelligent Automation & Document AI
Repetitive knowledge work handed to systems that read, extract, check and act, with a person in the loop where judgement matters.
A large share of back-office time goes on moving information from one place to another. Reading an invoice and keying it into the ERP. Checking a purchase order against a contract. Compiling the same weekly report from five sources. This work is slow, error-prone and rarely what anyone was hired to do.
We build automations that handle the full path: read the document or message, extract what matters, validate it against your rules and systems, then complete the action. Anything outside tolerance is routed to the right person with the context already gathered.
Every automated step is logged with its source. Any figure in any system can be traced back to the document, email or cell it came from, which is what makes automation acceptable to auditors as well as operators.
Where it applies
- Invoice, purchase order and contract processing
- Email and request triage with drafted responses
- Report generation and spreadsheet automation
- Approval and exception workflows
- AI agents that complete multi-step tasks across systems
What you get
- Workflows that run end to end, with a full audit trail
- Exceptions routed to the right person with context attached
- Hours saved, measured per process
Built with
- Large language models
- Document AI and OCR
- Workflow orchestration
- API and ERP integration
- Python
04 / Converse
Conversational AI & Copilots
Assistants that answer, guide and take action in plain language, over WhatsApp, voice, web chat or inside your own tools. In Sinhala and English.
The people who hold the ground truth in a South Asian operation often work away from a desk, and often in Sinhala. Asking them to open an app and fill in a form in English is why reports arrive late or not at all. We build assistants that meet them in the channel they already use, in the language they already speak.
Inside the business, copilots answer questions from your own policies, manuals, contracts and data rather than the open internet. Every answer cites its source, and anything the assistant is not confident about goes to a person.
Mixed-language input is treated as the normal case. A sentence in Sinhala with an English product name and a number in either language is exactly what real users send, so it is exactly what we design and test for.
Where it applies
- Customer service assistants on WhatsApp and web
- Voice reporting for field, site and sales teams
- Knowledge assistants over policies, manuals and SOPs
- Dealer and distributor support copilots
- Plain-language questions answered from your data
What you get
- Assistants grounded in your own documents, with sources cited
- Hand-off to a person when confidence is low
- Analytics on what people actually ask for
Built with
- Large language models
- Retrieval-augmented generation
- Speech-to-text (Sinhala and English)
- WhatsApp Business API
- Vector search
05 / See
Computer Vision
Cameras that report what they see. Inspection, counting and safety monitoring that runs on existing CCTV or purpose-fitted hardware.
Many operations already have cameras pointed at the things that matter: production lines, loading bays, site entrances, shelves. Almost none of that footage is watched. Computer vision turns it into a sensor that never looks away.
We train models on your own images, under your own lighting, at your own line speeds. Generic models tuned on stock photography fail quickly on a dusty site or a fast conveyor, so the data collection and labelling are part of the work rather than an afterthought.
Where connectivity is limited, models run at the edge on the camera or a small local device, sending alerts and summaries rather than streaming video. That keeps bandwidth low and footage on your premises.
Where it applies
- Visual quality inspection on production lines
- PPE and safety compliance monitoring
- Stock, pallet and vehicle counting
- Shelf availability and planogram compliance
- Damage detection for assets and deliveries
What you get
- Alerts in real time, not reports after the fact
- Dashboards of defects, incidents and trends
- Models tuned on your own images and conditions
Built with
- Object detection
- Image segmentation
- OpenCV
- Edge deployment
- Video analytics
06 / Build
Software, Data & Platform Engineering
The systems AI runs on. Web and mobile platforms, data pipelines, dashboards and integrations, built to pass enterprise security review.
A model is only useful once it reaches the people who make the decision. That means a portal they log into, a dashboard they check every morning, an integration that writes back into the ERP. We build those systems ourselves, so the AI and the software around it are designed together.
Underneath sits the data. Most AI projects stall because the data is spread across systems that do not talk to each other. We build the pipelines and warehouses that bring it together cleanly, which pays off well beyond any single model.
Enterprise systems have to survive a security review. We design identity, access control and audit logging first, and treat your control list as a specification rather than a hurdle at the end.
Where it applies
- Customer, supplier and dealer portals
- Internal tools and scheduling platforms
- Mobile apps for field and sales teams
- Data pipelines and warehouses
- Integrations with ERP, CRM and existing systems
What you get
- Production systems with role-based access and audit trails
- Documentation and handover your own team can maintain
- Deployment on your cloud or managed by us
Built with
- Next.js
- React Native
- Node.js
- Python
- PostgreSQL
- Cloud deployment
Industries
Built for the industries that run South Asia.
The models change from one sector to the next. The discipline does not: understand the operation first, then build for it.
FMCG & Food Manufacturing
- Demand and raw material price forecasting
- Production scheduling
- Line quality inspection
Construction & Engineering
- Voice-based site reporting
- Material tracking across sites
- Safety and PPE monitoring
Logistics & Distribution
- Route and load optimization
- Inventory placement
- Delivery exception handling
Retail & Supermarkets
- Store-level demand forecasting
- Shelf availability monitoring
- Customer assistants on WhatsApp
Banking & Finance
- Document and application processing
- Credit and churn risk scoring
- Customer service automation
Telecommunications
- Churn prediction
- Support ticket triage
- Field team scheduling
Manufacturing
- Visual defect detection
- Maintenance planning
- Procurement optimization
Healthcare & Public Sector
- Records digitization
- Resource and staff scheduling
- Citizen and patient assistants
Ways to Work With Us
Start small. Scale what works.
2 to 4 weeks
Diagnostic
Our engineers map your operation and return a ranked roadmap of AI opportunities, each sized by value, effort and data readiness. You keep the report whether or not you build with us.
Best when you know AI matters but not yet where to start.
6 to 12 weeks
Fixed-scope build
One system, scoped and priced before we start, built into production against live data. Payments are tied to milestones you can verify.
Best when the problem is clear and you want it solved.
Monthly
Managed AI service
We run the system after it ships: retraining as conditions change, monitoring for drift, and a dated monthly deliverable that reports its own accuracy against the last one.
Best for forecasts and models that must stay accurate for years, not weeks.
How We Work
From conversation to production.
- 01Consultation
Free · 30 to 45 min
We listen.
A structured conversation about your pain points, what you have tried and what success looks like. We share what we have seen in similar operations. No sales pitch.
We come prepared with questions, not slides. You describe the operation and where it hurts. We ask about volumes, decision cycles, who owns the number today and what has already been tried.
By the end of the call you get our honest view on whether AI is the right tool for the problem. Sometimes the answer is a process change and a better spreadsheet, and we will say so.
- 02Diagnostic
2 to 4 weeks
We map.
Our engineers spend time inside your operation through interviews, data pulls and process mapping. You get back a ranked roadmap of AI opportunities with ROI estimates, and you keep it either way.
Engineers sit with the people who do the work. We pull real data, trace how decisions are actually made and find where information is lost between systems.
The roadmap ranks each opportunity by value, effort and data readiness, with the assumptions written down. It is written so that anyone could build from it, including your own team.
- 03Pilot Build
6 to 12 weeks
We build.
One focused system from the roadmap goes into production. Fixed scope, fixed timeline, milestone-based payments. A working system in your hands, not a slide deck.
We take the highest-ranked opportunity and build it properly: validated, secured, documented and running against live data by the end of the engagement.
Scope and timeline are agreed before we start, and each payment is tied to a milestone you can verify. If your organisation runs a security review, its control list is a design input from the first day.
- 04Scale
Ongoing
We compound.
Pilots that work expand across sites, functions and regions. A managed service keeps them accurate as conditions change. A long-term partnership, not a handover.
A system that works in one plant usually works in five. Expansion is mostly integration and change management, which is faster and cheaper than the first build.
Our managed service keeps models accurate after launch: retraining, drift monitoring and a monthly report on real-world accuracy, long after the build team has moved on.
Questions
Answered honestly.
- What kinds of problems is G22 Neural best suited for?
- Problems where a recurring decision or a repetitive task costs real money when it goes wrong. That covers forecasting demand and prices, planning inventory and routes, processing documents, answering customers, inspecting products on camera, and building the platforms that tie these together. The common thread is an operation with real data and a measurable cost of error. We are a poor fit for research with no operational owner, and for problems a process change would solve more cheaply than software.
- How is domain-fluent AI different from generic AI?
- A generic team fits a model to your data. A domain-fluent team knows what belongs in the data. A commodity forecast moved from around 70% to 97% accuracy because we knew which weather index leads a South Asian harvest, not because of a more exotic algorithm. Knowing how a procurement cycle runs, what a planner actually needs to see and how a site really reports is the part that changes the result.
- Do you work with SMEs or only large enterprises?
- Both, through different engagements. Large enterprises usually start with a diagnostic and move to a pilot build. Smaller companies often start with a managed service, which is a much smaller commitment and produces value in the first month. What we cannot do well is an enterprise-sized build on a small budget, and we will tell you that in the first conversation rather than three months in.
- How does pricing work?
- The consultation is free. A diagnostic is a fixed fee based on the size of the operation being mapped, and you keep the report whether or not you build with us. Builds are fixed scope and fixed timeline, with payments tied to milestones, so you know the total before we start. Managed services are a monthly fee. We do not publish rates because they depend on scope and data readiness, but you always get a written number before any commitment.
- What technologies do you use?
- We choose well-supported tools that your team can maintain after handover. Python and gradient-boosted models for forecasting, mathematical optimization for planning, large language models with retrieval for assistants and document work, and object detection for computer vision. Applications are built in Next.js, React Native and PostgreSQL, deployed on infrastructure you approve, including your own cloud where policy requires it.
- How do you handle data security and PDPA compliance?
- We work under NDA from the first data access. Data is processed on infrastructure you approve, encrypted in transit and at rest, with access limited to the engineers on your engagement and every access logged. Under Sri Lanka's Personal Data Protection Act we act as a processor, not a controller: we hold personal data only where necessary and for as long as the engagement requires, and return or destroy it on request. If your security team maintains a control list, it becomes a design input at the start of the build.
- Do you sign non-disclosure and exclusivity agreements?
- NDAs, always, and usually before the first detailed conversation. Exclusivity is negotiable and scoped: we will agree not to build the same system for a direct competitor for a defined period in a defined market. We will not agree to blanket exclusivity across a whole industry, because that would stop us doing the work that makes us useful to you.
- What happens if a project doesn't deliver expected results?
- The structure is designed so you find out early and cheaply. The diagnostic tells you whether the data supports the result before anyone builds, and we will recommend against a build when it does not. Payments are tied to milestones you can verify, so a project that is not working can be stopped at a milestone rather than at the end. If a system we deliver misses the accuracy or performance targets written into the scope, we keep working on it at our cost until it meets them or you decide to stop.
- Do you offer support after project delivery?
- Yes, and we recommend it. A model handed over at the end of a project starts to drift as markets move and product mixes change. Our managed service covers retraining, monitoring, drift detection and a dated monthly report on accuracy. You can also take the system entirely in-house: everything is documented and the code is yours.
- How do you compare to global consultancies like Accenture or Deloitte?
- They bring scale, global benchmarks and the institutional weight some boards require. We are faster, considerably more affordable and closer to the ground. Our engineers work inside your operation rather than flying in for workshops, we build in Sinhala and English, and the people who scope your project are the people who build it. If you need a large transformation programme, use them. If you need a working system in production in twelve weeks, built by people who understand a South Asian operation, use us.
Not sure which of the six you need?
That is what the consultation is for. Thirty minutes, no cost, and an honest answer about whether AI is the right tool for your problem.