{ "schemaVersion": 1, "version": "2026.09.09", "id": "https://hk.onyxdevslab.com/data/case-study-evidence.json", "name": "Onyx Devs Lab enterprise AI case-study evidence register", "alternateName": [ "Onyx enterprise AI evidence dataset", "Onyx case-study metrics dataset" ], "publisher": { "name": "Onyx Devs Lab", "legalName": "ONYX DEVS LAB LIMITED", "businessRegistrationNumber": "79051925", "lei": "254900Z30CLK7HKE9H46", "url": "https://hk.onyxdevslab.com/" }, "dateModified": "2026-09-09", "landingPage": "https://hk.onyxdevslab.com/en/methodology/case-study-evidence-register/", "sameAs": "https://github.com/0xHunterL/onyx-devs-lab.github.io/releases/download/geo-evidence-2026-09-09/case-study-evidence.json", "keywords": [ "enterprise AI", "AI consulting", "custom AI development", "Forward Deployed Engineering", "case studies", "validation metrics", "Hong Kong" ], "evidenceClass": "Anonymised first-party project validation snapshots", "limitations": [ "Not independently audited", "Not a universal performance benchmark", "Client identities, raw records, and confidential baselines are not published" ], "methodologyUrl": "https://hk.onyxdevslab.com/en/methodology/case-study-evidence-register/", "cases": [ { "id": "retail", "name": "Turning retail transaction data into operating decisions", "sourceUrl": "https://hk.onyxdevslab.com/en/case-studies/retail-ai-decision-platform/", "operatingChallenge": "Transaction data existed, but inventory, workforce, and loss signals were fragmented. The business needed better decisions without replacing the operational ERP.", "deliveredIntervention": [ "Connected existing transaction data through a separate analytics layer.", "Designed missing workforce and inventory-loss data modules.", "Applied inventory forecasting and market-basket analysis to operational decisions.", "Kept recommendations reviewable by operators rather than automating high-impact decisions invisibly." ], "evidenceBoundary": "The published metrics come from project records and describe the delivered scope and pilot validation. They should not be extrapolated as expected results for another organisation.", "metrics": [ { "value": "51", "name": "stores covered", "definition": "Cross-store sales, inventory, and operating analysis" }, { "value": "289", "name": "legacy tables mapped", "definition": "Fields, sources, and metric ownership documented" }, { "value": "99.6%", "name": "reconciliation consistency", "definition": "Daily sales and inventory metrics checked against source systems" }, { "value": "77.2%", "name": "actionable inventory alerts", "definition": "Alerts followed by replenishment, transfer, or stock-reduction action" }, { "value": "21 min", "name": "median decision-data time", "definition": "From a cross-store question to a reviewable answer" } ] }, { "id": "accounting", "name": "An AI-native production line for accounting operations", "sourceUrl": "https://hk.onyxdevslab.com/en/case-studies/accounting-ai-production-platform/", "operatingChallenge": "Accounting work was split across client files, collection, bookkeeping, review, tax preparation, and exception handling. AI needed to assist the full workflow without obscuring evidence or professional accountability.", "deliveredIntervention": [ "Mapped 12 business domains, 52 screens, and 9 production workspaces.", "Separated deterministic accounting rules from AI judgement.", "Designed queues, retries, status tracking, evidence chains, and explicit human takeover.", "Kept final official submissions in a controlled path where direct APIs were unavailable." ], "evidenceBoundary": "The published figures separate mapped product scope from pilot workflow validation. They do not claim financial outcomes, and regulated decisions and official submissions remain subject to qualified human review.", "metrics": [ { "value": "12", "name": "business domains modelled", "definition": "Customer, invoice, bookkeeping, filing, and compliance operations" }, { "value": "52", "name": "screens and flows mapped", "definition": "Interface operations translated into orchestrated production steps" }, { "value": "9", "name": "production workspaces", "definition": "Batch work, exception takeover, and permission governance" }, { "value": "72.8%", "name": "low-risk task automation", "definition": "Tasks completed without rewriting and passed deterministic checks" }, { "value": "99.2%", "name": "evidence-chain completeness", "definition": "Critical judgements retain source, rule, version, and action record" }, { "value": "4.7 h", "name": "average month-end cycle", "definition": "From complete documents to a result ready for human review" } ] }, { "id": "legal", "name": "Legal AI that returns every conclusion to its evidence", "sourceUrl": "https://hk.onyxdevslab.com/en/case-studies/legal-ai-evidence-workflow/", "operatingChallenge": "Law-firm operations and large case files created two different information problems: fragmented business status and high-risk evidence retrieval across very large dossiers. A fluent answer was not enough; lawyers needed coverage, provenance, and explicit review.", "deliveredIntervention": [ "Connected leads, clients, matters, milestones, tasks, contracts, and receivables in one operating model.", "Tracked ingestion by file, page, and text chunk so missing or failed material remained visible.", "Combined keyword and vector retrieval with file, page, source text, and confidence coordinates.", "Applied matter-level access, audit trails, exception queues, and lawyer approval before formal output." ], "evidenceBoundary": "The validation figures come from a controlled project sample and lawyer-labelled reference set. They demonstrate the tested workflow, not legal accuracy for every jurisdiction or matter.", "metrics": [ { "value": "2", "name": "operating chains unified", "definition": "Law-firm operations and large-dossier analysis" }, { "value": "4", "name": "evidence coordinates", "definition": "File, page, source text, and confidence" }, { "value": "98.6%", "name": "citation-location accuracy", "definition": "Lawyer sample matched citations to the correct material and page" }, { "value": "92.4%", "name": "priority-material recall", "definition": "Lawyer-labelled key facts entered the candidate set" }, { "value": "11.8 h", "name": "case initial-review cycle", "definition": "Effective work time to the first issue list; 18.5 h before" } ] }, { "id": "recruiting", "name": "Recruiting automation with context, evidence, and send control", "sourceUrl": "https://hk.onyxdevslab.com/en/case-studies/recruiting-ai-agent-workflow/", "operatingChallenge": "Fast reply generation did not solve incomplete conversation context, stale candidate profiles, duplicate sends, or uncertain delivery. Automation needed to coexist with recruiters rather than take over their working interface.", "deliveredIntervention": [ "Used a local connector to reuse the authorised recruiter session and synchronise visible candidate history.", "Refreshed structured profiles when context changed, separating facts, inferences, confidence, and missing information.", "Supported manual, copilot, and controlled-autopilot modes per candidate.", "Serialised sends, enforced idempotency, and reread the sent message to verify completion." ], "evidenceBoundary": "The published validation covers synchronisation, send verification, and workflow time in a controlled pilot. It does not claim improved hiring outcomes without a longer recruiting cohort.", "metrics": [ { "value": "3", "name": "automation modes", "definition": "Manual, Copilot, and controlled Autopilot" }, { "value": "4", "name": "operating workspaces", "definition": "Candidate queue, conversation, profile, and operations" }, { "value": "98.9%", "name": "conversation-sync coverage", "definition": "Visible historical messages successfully backfilled" }, { "value": "99.3%", "name": "send reread success", "definition": "Sent text reread from the recruiter side after action" }, { "value": "15.3 min", "name": "candidate handling time", "definition": "First review to confirmed next action; 22.4 min before" } ] }, { "id": "industrial", "name": "A trusted data path from industrial operations to management decisions", "sourceUrl": "https://hk.onyxdevslab.com/en/case-studies/industrial-erp-ai-data-platform/", "operatingChallenge": "Production information was fragmented across equipment, paper records, and isolated systems. Management received delayed summaries, but connecting the data could not compromise safety-related industrial control.", "deliveredIntervention": [ "Mapped material movement, weighing, quality testing, vehicle energy, and production records to accountable business objects.", "Established a controlled data interface between PLC information and management systems.", "Unified operational dashboards, mobile access, permissions, validation, and audit records.", "Added business questions and exception explanation above the trusted data layer, not direct AI control of production." ], "evidenceBoundary": "Validation covers data-record completeness, collection timeliness, and operator acknowledgement in the delivered project scope. Safety-related actions remain under authorised site personnel.", "metrics": [ { "value": "4", "name": "field processes connected", "definition": "Weighbridge, laboratory, vehicle, and energy" }, { "value": "1", "name": "unified operating entry", "definition": "ERP, mobile access, and industrial data" }, { "value": "96.1%", "name": "field-data coverage", "definition": "Critical shifts, equipment, and business events recorded" }, { "value": "99.2%", "name": "on-time collection", "definition": "Data arrived within the defined dashboard window" }, { "value": "27 min", "name": "exception acknowledgement", "definition": "Alert to accountable operator confirmation; 41 min before" } ] }, { "id": "credit", "name": "Credit research automation without hiding the source", "sourceUrl": "https://hk.onyxdevslab.com/en/case-studies/credit-research-ai-agent/", "operatingChallenge": "Analysts repeatedly collected, reconciled, and reformatted information across internal files and public sources. Faster drafting was useful only if material facts, conflicts, freshness, and provenance remained visible.", "deliveredIntervention": [ "Modelled the research task, evidence levels, report structure, and analyst approval boundary.", "Connected internal materials with current public financial information through a deduplicated retrieval pipeline.", "Separated facts, inference, and risk judgement while attaching sources to material conclusions.", "Recorded analyst edits and missing evidence as evaluation data for future releases." ], "evidenceBoundary": "The validation sample measures fact recall, citation coverage, and first-draft time. The system does not replace credit, investment, or trading decisions.", "metrics": [ { "value": "2", "name": "research source classes", "definition": "Internal material and current public financial information" }, { "value": "3", "name": "orchestrated stages", "definition": "Retrieval, cross-checking, and report drafting" }, { "value": "91.6%", "name": "priority-fact recall", "definition": "Reference facts entered the evidence set" }, { "value": "98.3%", "name": "conclusion citation coverage", "definition": "Material report judgements retained a reviewable source" }, { "value": "4.4 h", "name": "research first-draft cycle", "definition": "Task creation to analyst-ready draft; 7.6 h before" } ] } ] }