Case studies.

Cross-industry · OPNYX productShipping

AgenTorQ — governed AI agents on Salesforce

Challenge. Enterprises wanted AI inside Salesforce, but could not accept a layer that bypassed sharing rules, metered unpredictably, or left no audit trail.

A managed AppExchange package with three deployment shapes — inside the customer's own org, a cloud workspace, or self-hosted. Personal data masked before the model sees it, writes gated behind an approval card, everything written to an immutable audit log, and enforcement in USER_MODE so sharing and field-level security hold inside the AI layer.

5Packages shipped

563Apex classes

10In-place upgrades

0Flows shipped

Capital markets · OPNYX productComing soon

FinTorq — markets screener and research terminal

Challenge. Retail forecasting tools advertise accuracy figures that do not survive out-of-sample testing.

A screener across 456 verified instruments in ten markets with three walk-forward models. The product states plainly that direction is not forecastable — on the benchmark index it scores 69.8%, and so does a naive always-up rule. Volatility regime, which is forecastable, is what shipped.

456Instruments verified

74%Volatility regime called

9sFull US scan, cold

Public procurement · OPNYX productComing soon

BidTorq — public tender aggregator

Challenge. Tender aggregators recycle stale caches, and one major public API silently ignores every keyword parameter it accepts.

Four live public sources with a real BM25 inverted index, a query language covering country, value, deadline, phrases and negation, and deterministic fit scoring driven by the customer's own profile keywords. The API that ignores its own keyword parameters is filtered locally rather than trusted.

4Live sources

54/54Browser assertions

362Live notices verified

Cross-industry · internal R&DFeeding the product

AgenTorQ research platform

Challenge. Capabilities proven in a demo often break in a packaged build, so they need somewhere to be broken first.

Dedicated research orgs carrying an intent and directive framework — intents declared as records, chained to reusable directives, dispatched through a handler registry resolved by type name. Eight handler implementations, a studio interface for composing intents without code, a universal file processor normalising documents for thirteen model providers, and a knowledge bridge ranking chunks by similarity computed in Apex.

24Model engines carried

229Apex classes

8Directive handlers

B2B commerceReleased 1.0.0.1

B2B Migration Accelerator

Challenge. Moving a live commerce estate onto Salesforce meant four different source platforms and a mapping that no two customers would agree on.

A managed AppExchange package pulling whole source records from Shopify, BigCommerce, CloudCraze and commercetools, with an inline mapping editor covering any createable target object, per-field transforms, assisted auto-mapping, and pause-and-resume mid-migration. Three earlier versions failed in the field on credential architecture alone — each cause is documented in the install guide.

4Source platforms

31/31Apex tests

0Hardcoded mappings

ManufacturingReleased v0.1.0-4

Agent Accelerator — an agent as a package

Challenge. An Agentforce agent is normally rebuilt per org. The client needed one that installs from a link and can be extended by the customer.

A managed, namespaced package shipping the entire agent as a reusable bot template — the shell, its planner bundle, a 56-tool metadata registry and an extensibility kit with a sample tool and a written procedure for adding one. Two component types cannot be packaged under platform limits; the resulting five-minute per-org setup is documented rather than hidden.

199Files validated

173Tests passing

56Leaf tools

ManufacturingPackaged 0.1.0.2

Manufacturing Agentforce agent

Challenge. Every new agent capability meant editing the planner and redeploying the dispatcher — slow, and risky in a live org.

A production agent whose entire tool surface is a custom metadata registry, so adding a capability is three files. Packaging required solving three constraints: the native AI class is not reliably provisioned in a fresh scratch org, commerce schema types were compile-time dependencies, and coverage collapsed in a bare org because the scratch administrator holds no field-level security on the package's own fields.

36Agent tools

86%Coverage, from 55%

15Findings, from 165

Cross-industryDelivered

Data AI Readiness accelerator

Challenge. Clients asked whether their data was ready for AI, and wanted an answer backed by their own records rather than a generic maturity model.

Apex only samples and measures; the model performs the statistics, scoring and confidence, grounded in and citing the real sample — heavy statistics in Apex would meet governor limits and produce untraceable figures. Includes governed enrichment write-back where the model proposes per-record corrections and Apex applies them with object and field-level security enforced.

200Record write-back cap

16Apex tests

100%Runs persisted

B2B commerceDelivered

Cart Sync framework

Challenge. Cart-to-quote synchronisation was being rewritten for every object pairing a customer asked for.

An administrator console discovers eligible objects through a curated seed list and a dynamic schema scan, stores a header and line pairing with its field mapping, and a sync engine writes to whichever targets are enabled. No object name appears in the engine. Two components are the only points a user touches; the configuration object stays backend storage.

0Object names in engine

2User touchpoints

3Apex services

Public sector, regulatoryDelivered

Regulatory fees, summonses and payments

Challenge. A regulatory-fee field rename across a live platform, where authorisation covered one environment out of five and a published data dictionary had to stay accurate.

The rename propagated across an Apex controller, five record pages, two Flows, four layouts, seventeen history-tracked fields and the system administrator permission set — with a manual runbook written for the environments the team was not authorised to touch. Also a code-mapping metadata type with ten records and an autolaunched sub-flow branching across summons and seven settlement categories.

92Files in one story

47Dictionary rows closed

27Stale permissions removed

Software & SaaSLive in testing

CPQ quoting and deal desk

Challenge. Sellers were clicking through managed-package screens to configure a quote, and the deal desk was reading every record by hand.

A single-page quote configurator on both Opportunity and Quote with tiered price allocation, typeahead search, staged deletions with undo and roll-up to the opportunity. Alongside it, a multi-object AI hub triaging records into three lanes, and a comparison scorecard written by two independent writers so a record scores itself the moment it is created.

97%Controller coverage

304/304Components validated

82/82Scorecard tests

Building materialsWith the client

Fixed-bid CRM programme, 15 entities

Challenge. A North American group of fifteen operating entities needed a fixed price, which meant every open question had to carry effort rather than become an exclusion.

A statement of work covering Sales Cloud Enterprise, Data 360 and Marketing Cloud Advanced, with licensing named precisely against the actual order rather than an earlier assumption. The holiday period was costed as reduced throughput rather than ignored, and the roadmap deliberately simplified to ten phase bars because a client reads a plan they can hold in their head.

1,009Person-days

22Weeks

12Payment milestones

Consumer goodsDelivered

Field execution cockpit, offline-first

Challenge. Field representatives cover eight to twelve stores a day, often without reliable connectivity, and needed AI insight grounded in what is actually on the shelf.

A regional cockpit covering shelf inventory, promotions, surveys and assisted ordering, with an offline queue carrying local and synced state indicators and a replayable audit payload. Seed data carries deliberate imperfections — a zeroed facing, missing signage, a declining trend — so the recommendations answer something real.

8Custom objects

7Standard-object extensions

100%Offline replayable

Legal servicesTwo orgs in sync

Full Data 360 build for a claimant firm

Challenge. The brief arrived as thirty-one populated spreadsheet cells and four diagrams, for a multi-practice firm mid-way through replacing its case-management platform.

The team read the client's actual schema — 68 objects, roughly 1,100 fields — and built the Data 360 layer designed to survive the migration. Party is modelled on Account rather than Contact with flat phone and email fields, so contact-point data model objects needed an unpivot rather than a mapping. The most useful finding was that the firm has no legacy package traces at all: they had already built on the standard model.

46Objects built

7,381Rows seeded

25–40mFull rebuild, any org

Vendor managementClosed

Data Cloud enablement — the correct diagnosis

Challenge. An org could not be enabled for Data Cloud. Licences were present and configuration looked correct, so the assumption was a missing setting.

The preference was observed reverting within minutes, applied by an automated process in the audit trail — a signature meaning the org has no entitlement. The conclusion, that this needed a support case rather than another toggle, ended a search that would otherwise have run for days, and a working laboratory org was stood up in parallel so the build could proceed regardless.

1Decisive endpoint

2Orgs A/B compared

1Own claim withdrawn

Industrial manufacturingLive

Executive quality twin

Challenge. A console aimed at a chief executive, where any figure that drifted from the pipeline would destroy its credibility in the room.

The entire warehouse ships to the browser and the same view definitions execute there — one file, two engines, no opportunity to diverge. Shipping a fifteen-table subset had left six consoles on stale extracts; the whole warehouse now ships. Verified by a harness that loads the site twice, once with the warehouse blocked at the network so the engine cannot start, and compares every number-bearing text node.

39/39Datasets by SQL

41Tables shipped

0Hand-authored figures

Pipe manufacturingLive

Production scheduling with a real solver

Challenge. Scheduling demonstrations narrate their what-if scenarios. This audience schedules for a living and would see through that immediately.

A finite-capacity constraint engine with resource calendars, per-plant shift offsets, simultaneous machine and crew booking, sequence-dependent changeover from a matrix, fixed cure windows and five dispatch rules. One engine, two call sites: it builds the warehouse and re-runs in the browser tab, then asserts on screen that the default rule reproduces the published figures.

81.7%On-time delivery

669Operations placed

1393%Peak MRP capacity ask

TelecommunicationsDelivered

Servicing console for a case-management RFI

Challenge. An RFI panel needed to see a working console on their own instance, not a mock bound to invented data.

An installable update-set package that binds to the target instance's own incidents, cases and knowledge — so it ships no demonstration data by design, and the accompanying deck says so. A 27-slide evidence pack generated from a single content model into both presentation and print formats, with screenshots captured live from the running instance.

366Package updates

~5mInstall time

19API sweep assertions

TelecommunicationsLive

Service-health console with a live simulator

Challenge. The intended monitoring module was not in the instance's plugin catalogue and its alert table returned an error, so the planned demonstration was impossible.

Rather than simulate it in the interface, the console was built on a custom scoped application and driven from an external estate simulator over the REST API — so the integration shown is genuine. The handover documents the platform's silent-failure modes, each of which fails without raising an error. The absence of real devices behind it was published rather than concealed.

13Deploy scripts

2Surfaces, one API

1Limitation published

Cross-industryProduct build

Multi-tenant service console

Challenge. A console that had to be sold to more than one customer, and built by several engineers in parallel without collision.

A 763-line contract document as the single source of truth — data model, schema, repository API, event bus, module registry, design system and every component property, with signatures declared stable. Module isolation is a firm rule, the data layer is provider-swappable so feature modules need no change when the backend lands, and retheming is a buyer action rather than a fork.

763Contract lines

7Dependency allowlist

3Demo roles

Retail marketplaceArchitecture agreed

Marketplace order and inventory orchestration

Challenge. Marketplace orders were being fulfilled by retail stores keying them in by hand, with no central view of which orders came from which platform.

Store-level availability, sourcing and a control tower rather than a warehouse system — the manual keying is the thing being removed. Postgres-only with two services, queueing and search held in the database because four services would not meet the five-minute deployment requirement. Multi-tenant with row-level security and the first tenant seeded from configuration rather than code.

17Country-platform pairs

8Brands

2Services to deploy

Logistics & haulageBid submitted

Competitive bid — measured, not assumed

Challenge. Four firms bidding, three recommending a full rebuild, and a client explicit that he was assessing business understanding rather than technology.

All eight supplied accounts were signed in to and driven through eighteen dashboard pages each under browser automation, against the live prototype. Role-based access control was found to work as intended — both driver accounts redirected away from every restricted page, eight blocked attempts each — so the permission model carries over instead of being rebuilt, which reduced the estimate. One earlier finding was withdrawn in writing after re-testing.

144Page loads driven

8Accounts swept

1Finding withdrawn

Network operationsPlan published

Off-platform network management assessment

Challenge. The client put six open-source projects forward as candidate bases and asked which to build on.

Each scored on capability, licence and architecture. Only one is a genuine network management system, and its copyleft licence and C++ codebase place it behind a REST boundary as an optional service rather than as the base. The recommendation: own the platform and console, take device input and output from permissively licensed libraries, and make the environment real with a container-based network lab plus a simulator for scale.

6Candidates scored

8Phase plan

1Licence table published

Public sector patternDelivered

Identity validation and third-party integration

Challenge. A compact reference build showing how a small feature is written end to end, from validation through to a credentialled third-party call.

A national identity number validated by checksum, date and citizenship digit — in the browser and again on the server, so the client-side check is convenience and the server-side check is the rule. Persistence keyed on a unique external identifier with a search counter, so re-submitting the same identity updates rather than duplicates. The third-party API sits behind a Named Credential with a placeholder committed to source.

18Tests, mocked callouts

95%Decoder coverage

2Validation layers