13 domains · 7 with shipped systems behind them
Most vendor industry pages read identically because they are lists of domains with the adjectives swapped. This one marks each domain with whether a system exists behind the claim, then describes the actual data problem rather than the trend.
The recurring shape across all of them: the volume is larger than the business thinks, the accuracy problem is upstream of the model, and the return comes from removing manual reconciliation rather than from adding a dashboard.
Grade control, fleet telemetry, plant availability
Mining decisions are capital-heavy and mostly irreversible. Whether a load is ore or waste is decided in seconds at the shovel, and material sent to the waste dump does not come back. The block model, the mine plan and what the mill actually received are three different numbers, and the gap between them is where recovered value leaks.
Capability. We have not shipped a mining system. What transfers directly is high-frequency telemetry ingestion, out-of-order and duplicate event handling, and running systems that cannot be paused for a maintenance window.
Fleet management telemetry from haul trucks and drills at sub-second intervals, fixed-plant sensors on conveyors and mills, survey point clouds from drone and LiDAR runs, weightometer readings, assay results from the lab, and dispatch and ERP records on top. Retention horizons are long because reconciliation is annual, not weekly.
Pit radio coverage is intermittent, so device clocks drift and events arrive out of order or twice — a truck that logs one load as two inflates mined tonnes for the whole shift. Weightometer calibration drift biases every downstream tonne silently. Assays land days after the material has moved, so the model and the plant disagree by design and the pipeline has to represent that rather than paper over it.
Short-interval grade prediction at the shovel, anomaly detection on mill and conveyor vibration and motor-current signatures, haul-cycle time prediction to feed dispatch. Not a replacement for the resource model or for a geologist signing off on a reconciliation.
Two mechanisms, both measurable on site: fewer misrouted loads, and fixed-plant degradation caught before it becomes an unplanned outage. Correctly attributed downtime is what makes an availability figure worth acting on.
Pricing, claims, exposure and reserving
Every policy is a priced position on a loss distribution. Pricing accuracy and reserving both depend on data the insurer does not fully control — address quality, hazard mapping, cause-of-loss coding entered by whoever took the call.
Capability. Adjacent to delivered risk, pricing and settlement work, but we have no insurance delivery reference and will not present one.
Quote traffic dwarfs bound policies, so the pipeline sizing is driven by quotes rather than sales. Add claims document and image packs, telematics for motor books, geospatial hazard layers for flood, bushfire and cyclone exposure, and reinsurance treaty data.
Geocoding precision decides which hazard zone an address sits in and therefore what the premium should be, so a mis-geocoded address is a mispriced policy that nobody notices until a catastrophe event. Party matching across policies drives duplicate and linked-claim detection. Inconsistent cause-of-loss coding makes portfolio analysis unreliable no matter how good the model on top is.
Extraction from claims document packs, first-notification-of-loss classification, and fraud and triage signals. With a hard constraint: a declined or repriced decision has to be explainable and reproducible, so the model output feeds a rule that a human owns rather than becoming the decision.
Claims cycle time and claims leakage. Automating the low-complexity tail of claims is what frees assessors for the complex ones, which is where the recoverable money is.
Video analytics, civic sentiment, registries
In government the record is the service. A citizen has whatever entitlement the record says they have, so data quality is not an operational metric, it is the thing being delivered.
Delivered. Real-time individual tiger identification from live video for a state government wildlife authority in India in 2018, across a national park — 96% identification accuracy against stripe patterns, and roughly a 70% reduction in manual surveillance workload. Separately, a civic sentiment analytics platform for a government end user in 2018–19, ingesting news, citizen complaints, feedback forms and automated phone surveys, at 92% classification accuracy. Both were delivered via an enterprise partner. Where an engagement sits on a statutory decision path we would still expect to work alongside a partner who has been through that procurement and assurance process.
Registry and transaction systems with decades of history, asset and sensor data for infrastructure, correspondence at scale, and open-data publication obligations on top of all of it.
The schema is a legal artefact: a field means what the legislation says it means, and when the legislation changes the old values stay valid for the period they covered. That makes effective-dated records mandatory rather than a nice-to-have — you cannot overwrite. Identity resolution without a single identifier produces duplicates. Corrections must be auditable and never silent, because someone will eventually have to explain a determination.
Classification and routing of correspondence, and search across policy and case corpora. Explicitly not automated determinations that affect an entitlement — those need a human decision-maker and a reviewable trail, and building it any other way creates a problem no amount of accuracy fixes.
Usually avoided cost rather than revenue: fewer manual re-keys between systems that cannot be decommissioned, and audit and information-request responses assembled from a queryable store instead of by hand. Accessibility obligations and data residency are requirements with test cases, not a compliance appendix.
Fixed odds, tote pools, exotics settlement
The price is the product. It is derived from form, scratchings, track condition and pool movement, and it has to be correct at the moment a bet is struck rather than shortly afterwards. Everything downstream — risk, liability, settlement — inherits whatever the price got wrong.
Delivered. Ashva, our multi-tenant tote and fixed-odds betting platform, runs in production for a tier-one wagering operator: per-bet routing to external tote hosts and fixed-odds operators, in-house risk retention, fixed-odds pricing, settlement, wallets, promotions and per-tenant real-time reporting from one deployment. We have delivery history with Australian wagering clients, and the operations desk covers Australian racing hours.
Load is concentrated, not averaged. Saturday metropolitan meetings and the spring carnival set the ceiling, and within a race a large share of turnover arrives in the last ninety seconds before the jump. Every price update fans out to every connected client, so the write volume is small and the fan-out is enormous.
Feed providers disagree with each other on scratchings, barrier changes and results, so the platform needs a resolution policy rather than a preferred provider. A withdrawal recognised at the wrong moment misprices every open bet on the race. Protests and amended placings force resettlement of bets already paid out, which means settlement has to be re-runnable rather than one-shot. A duplicated bet message that is not idempotent becomes a real liability on the balance sheet.
Pricing assistance, form feature engineering, and anomaly detection over betting patterns. Not unsupervised customer limit-setting, and not a model with authority to move a market without a trader in the loop.
Settlement that completes without manual intervention, including on protest days, and a trading desk that spends its time trading rather than reconciling two feeds that disagree. Both convert directly into headcount that scales with the book rather than with the number of meetings.
Wallets, content aggregation, responsible gambling
An operator running sports betting and casino content together has one commercial reality and two technical worlds. The wallet is where they meet, and it has to be the single source of truth or you get balance disputes you cannot resolve after the fact.
Delivered, scoped precisely: the shipped work is wallet, content integration and responsible-gambling enforcement inside a wagering platform. We have not built a standalone casino operator from scratch and do not claim it.
Every game provider ships its own session and round protocol, so integration count drives complexity more than traffic does. Round-level events for every spin and hand have to be retained for reconciliation and for return-to-player verification.
Round reconciliation against each provider ledger is the only way to find a divergence before a customer does. Currency handling and rounding at round level compound quickly across millions of rounds. Deposit limits, cool-off and self-exclusion have to be enforced at the wallet, because enforcing them in the interface means the next client application reintroduces the gap.
Play-pattern signals for responsible-gambling intervention, and content recommendation. The signal raises an intervention for a human to act on; the hard limits stay rules, because a regulator will ask you to show the rule.
Adding a content provider should be configuration rather than a project. That is the whole efficiency argument, and it is measurable in weeks per integration.
Order entry, options pricing, real-time risk
Market data is both the input and part of the definition of correctness — a price that arrives late is not merely slow, it is wrong. Position and risk have to be computed continuously rather than at the end of the day, because that is when the decision gets made.
Delivered. AnkEDGE, our options trading platform for options traders, is built and operated by us. Risk and settlement work in the wagering platforms draws on the same engineering.
Full-depth order book updates, options chains where a single underlying tick reprices the whole surface, and tick history retained for backtesting. Message rates are orders of magnitude above trade rates.
A sequence gap in a market data feed corrupts the book silently, so gap detection and snapshot recovery are mandatory rather than defensive extras. Corporate actions retro-adjust history, which means a backtest run today against unadjusted history quietly disagrees with one run last month. Clock discipline is an audit requirement, not a tidiness preference.
Signal research, execution-cost and slippage prediction, and surveillance over the firm own flow. Not a model holding position authority without a human gate — the failure mode is fast and expensive.
Fewer end-of-day breaks to chase, and the ability to replay a full trading session to settle a dispute in minutes instead of reconstructing it from logs over days.
Matching, market data fan-out, surveillance
An exchange sells fairness and determinism. Participants have to believe the matching engine treats them identically, and the operator has to be able to prove it after the fact.
Capability. Closely adjacent to the trading platform work, but we have not built or operated a licensed exchange and will not imply otherwise.
Order, cancel and replace traffic far exceeds executed trades, and market data fans out to every participant simultaneously. Every inbound message needs journalling because replay is the product of record.
Deterministic ordering and a replayable journal are the whole foundation: an exchange that cannot reproduce a session from its journal cannot defend an outcome to a participant or a regulator. Sequencing has to be fair, and timing jitter that leaks queue position to some participants and not others is a fairness defect even when the averages look fine.
Surveillance for wash trading, spoofing and layering, generating candidates for a human surveillance analyst to work. Not automated enforcement, and not a black box, because every alert may have to be explained.
Capacity headroom without proportional hardware, and a surveillance function that does not need headcount growth matched to volume growth.
Chain indexing, reconciliation, AML screening
Chain data is public but almost never directly usable. Turning it into something an operations or compliance team can rely on means reorg-aware indexing, and the auditability that makes the asset class interesting is exactly what makes the data engineering unforgiving.
Capability. No shipped crypto reference. We hold no financial services or digital asset licence of any kind and do not custody assets.
Multi-chain indexing with per-block state diffs, mempool observation, and full history retained rather than aged out, because the premise of the whole system is that history is verifiable.
Chain reorganisations mean a record you treated as confirmed can be un-confirmed, so an indexer that assumes finality is wrong rather than merely optimistic. Deposit crediting needs a confirmation policy per asset. Decimal precision and dust handling break naive ledger arithmetic, and address clustering is probabilistic and must be represented as such.
Flow anomaly detection to generate AML screening candidates. The sanctions and reporting obligations themselves stay rules, because a compliance officer has to be able to point at the rule that fired.
Unattended reconciliation between on-chain state and the internal ledger. That reconciliation is where the cost sits today, and it is the only line worth automating first.
Point of sale, inventory, pricing, loyalty
The till is the only place where price, stock and the customer meet at the same instant. Whatever is wrong upstream becomes visible there first, usually to a queue of people.
Delivered. AmshPOS, our point-of-sale system for retail operations, is built and operated by us.
Line-item transactions per store per day, price and promotion changes pushed to every lane, stocktake counts, and loyalty events across channels that all have to land in one customer view.
Tills have to keep selling when the link to head office drops, which means conflicting writes on reconnect and therefore a deterministic merge with idempotent transaction identifiers — get that wrong and you double-count a day of sales. Stock-on-hand drifts between store and warehouse until a count reconciles it. Promotion stacking produces prices nobody intended, so the pricing engine needs to be testable rather than configurable-and-hope. Receipt and tax rules vary by jurisdiction.
Per-SKU per-store demand forecasting, replenishment suggestions, and shrinkage anomaly detection at lane level. Useful precisely because the underlying transaction data is clean and complete, which is the part most retailers do not have yet.
Fewer manual stock adjustments, and less markdown on lines that were over-ordered because the forecast ran on head-office averages rather than store-level demand.
Monitoring, incident response, release management
Here telemetry is the product. An estate you cannot observe cannot be operated, and the quality of the data decides whether an on-call engineer is diagnosing or guessing.
Delivered. The desk runs our own platforms and client estates, 24×7 across time zones. We are in Ahmedabad, and Indian Standard Time puts our working day across Australian business hours — including Saturday metropolitan meetings and the spring carnival. There is no Australian office, entity or phone number behind that; the coverage comes from where we already sit.
Metrics, logs and traces from every service in the estate, retained long enough to investigate a pattern across weeks rather than only the incident in front of you.
Alert precision is the single metric that decides whether a desk works. An alert that fires on a symptom instead of a cause trains the roster to ignore it, and once that happens the monitoring is worse than none because it manufactures false confidence. Runbooks have to be versioned with the service they describe or they go stale silently.
Log clustering and first-pass triage to cut noise, and correlation of an incident to the release that caused it. Not automated remediation on anything with financial consequence without a human gate.
Fewer escalations reaching the engineering team, and release cadence maintained through peak periods instead of frozen through them. The team that built the system is the team on the pager, so there is no handover to a vendor who has never read the code.
Tiered support, defect triage, tenant escalation
Every ticket is a data point about a defect that has not been fixed yet. Treated as a queue to be drained, support generates no information; structured properly it is the highest-volume source of truth about how the product actually behaves.
Delivered, for the platforms we build and operate. Support and engineering sit in the same organisation, which is the reason a tenant escalation can reach the person who wrote the code.
Ticket history with session context, correlation identifiers and build versions, all of it needing to join against release and telemetry data to be worth anything.
A ticket without the tenant, the build version and a correlation identifier is not reproducible, so intake structure determines whether the next tier can act at all. Duplicate detection against known issues has to be reliable or the same defect is investigated repeatedly by different people.
Duplicate detection and suggested responses drawn from documented resolutions, with the answer traceable to a specific known issue rather than generated prose. A confidently wrong support answer costs more than a slow correct one.
First-contact resolution, and defects flowing back into the engineering backlog instead of accumulating as recurring tickets that quietly set the support headcount.
Imaging pipelines, records integration, triage
A diagnosis is a decision made on incomplete data with a hard consequence. That places the burden on completeness and provenance rather than on volume — knowing what is missing matters as much as what is present.
Capability. No healthcare delivery reference. We would not take a first healthcare engagement as sole vendor on a clinical decision path, and would say so in the first meeting rather than the third.
Imaging is the driver: DICOM studies are large and arrive in bursts from modalities rather than evenly. Pathology results, device telemetry and referral correspondence sit alongside it.
HL7 v2 is implemented differently at every site, so standards-based integration still means per-site mapping, and FHIR reduces that work without removing it. Patient identity matching without a single national identifier produces duplicate records that then diverge. Units and reference ranges differ between labs, so a result compared across providers without normalisation is misleading rather than merely noisy.
Triage prioritisation and imaging pre-read, as a second reader that flags for a clinician — never as the reporting clinician. Consent has to be enforced as a data-access control, not recorded as a checkbox next to data that is already readable.
Reporting turnaround time, and clinician hours currently spent reconciling records between systems.
Plant telemetry, traceability, downtime attribution
Two questions carry the value: why the line stopped, and which units a recall actually covers. Both are answerable only if per-unit genealogy and downtime cause were captured at the time, because neither can be reconstructed later.
Capability. No manufacturing delivery reference. The transferable part is high-frequency telemetry ingestion and 24×7 operation of systems that do not stop.
PLC and SCADA tags at high frequency over OPC-UA, per-unit genealogy records, and quality inspection images from every station that has a camera.
Device clocks drift and plant network links drop, so edge buffering plus out-of-order tolerant ingestion is mandatory. A gap-filled series that looks continuous is more dangerous than a visible gap, because someone will trend it. Tag naming conventions differ per line, which makes cross-line comparison meaningless until a semantic layer maps them.
Predictive maintenance from vibration and motor-current signatures, and visual inspection. Said plainly: predictive maintenance needs enough labelled failure events to be worth the build, and most plants do not have them yet — the honest first project is usually instrumentation and downtime attribution, not a model.
Overall equipment effectiveness improved by attributing downtime correctly rather than to a default reason code, and recall scope narrowed by genealogy that holds up to scrutiny.
The pattern travels further than the vertical does. If the problem is high-volume ingestion, data you cannot trust, or a platform nobody wants to change, the conversation is the same one.
