{"resourceId":"ukraine-epermit-gemma-onprem-veterinary-license-review-20261106","versions":[{"version":"external-2552f12a261502c19a7ba85d16785513ed91b56ff1decdbe3942060e46f07bf6","resource":{"id":"ukraine-epermit-gemma-onprem-veterinary-license-review-20261106","title":"Ukraine Ministry integrates on-prem Gemma 3 27B into ePermit licensing review for veterinary licenses, recommending issue, reject, or manual review while officials retain decisions","organization":"Ministry of Economy, Ecology and Agriculture of Ukraine (Google DeepMind case study)","sector":"National digital permitting and licensing portal","geography":"Ukraine","publishedAt":"2026-11-06","publicationDate":null,"eventDate":null,"sourceName":"Google DeepMind Gemmaverse","sourceLabel":"Google DeepMind: The Ministry of Economy, Ecology and Agriculture of Ukraine digitizes licensing process with Gemma","sourceUrl":"https://deepmind.google/models/gemma/gemmaverse/epermit/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["developers-agents","data-security","governance-procurement"],"finding":"A Google DeepMind Gemmaverse case study describes Ukraine's ePermit service on the Diia portal integrating Gemma 3 27B for license application review after Qwen 2.5 7B VL extracts text from scanned documents and registry data are combined. Legal constraints require closed infrastructure with no external API calls or cloud dependencies; the model runs on-premise via vLLM on NVIDIA H100 GPUs in a Ukrainian datacenter. Gemma recommends one of three outcomes to the assigned official—issue the license, reject, or further manual review—for veterinary licenses initially, with expansion to other types planned. Officials quoted say Gemma balanced reasoning and on-prem deployment; the write-up claims dramatically cut processing time, no latency issues, and fewer rejections/resubmissions/workload, without publishing numeric baselines. Decisions are described as auditable via logged inputs, outputs, and rules.","sledRelevance":"Licensing/permitting AI with an explicit on-prem, no-external-API constraint and human final decision—transferable operating constraint for SLED agencies that cannot send applicant documents to public clouds.","evidence":"Vendor case study with named ministry CDTO/AI advisor quotes; impact claims lack numeric baselines or independent evaluation; retrieval date used as publishedAt because the page is undated.","architectureImplications":"Vision LLM for scan OCR plus Gemma 3 27B decisioning on-prem with registry joins; three-way recommendation to a human official; multimodal fine-tuning planned.","governanceImplications":"Human official retains decision; auditable logs of inputs/outputs/rules; evaluation framework planned as types expand.","securityPrivacyImplications":"Closed infrastructure mandate—no external APIs or cloud—addresses sovereignty; applicant document handling still depends on ministry controls not fully specified here.","caveats":"DeepMind marketing page; 'dramatically' faster processing without numbers; veterinary-only production scope; English typos in ministry name on the page.","streamIds":["citizen-engagement"],"roles":{"sales":"National licensing portals want AI review without sending applicant packets to public clouds. Ukraine's DeepMind case study shows on-prem Gemma recommendations with humans deciding, but speed claims lack numbers. Discovery should ask closed-infrastructure mandates, which license types are in scope, and how overrides are logged. A bounded engagement is a shadow-mode agreement rate on a labeled veterinary set. Do not quote 'dramatically faster' as a measured KPI from this page.","engineering":"Engineering fit is on-prem multimodal extraction plus LLM eligibility reasoning with a three-way recommendation to officials. Confirm air-gapped GPU hosting, registry connectors, complete audit logs, and that the model cannot auto-issue licenses without a human. Proof of value is shadow-mode precision and recall versus official decisions and time-to-first-feedback versus a published baseline—not vendor adjectives about dramatic speedups.","delivery":"Delivery work is license-type rule and evaluation sets plus official training on reading recommendations, not removing human decision rights. Ministry CDTO and IT own model operations while program offices own eligibility rules for each license type. Acceptance criteria below are proposals. Keep auto-issue off until evaluation gates pass, and log every override with a reason code."},"roleDetails":{"sales":{"customerProblem":"License applicants wait days for incompleteness findings while officials grind document checks by hand.","stakeholders":["Ministry CDTO / AI advisors","Licensing officials","Applicants","Datacenter/security teams"],"discoveryQuestions":["Which license types allow AI recommendations?","What is the closed-infrastructure boundary?","How are overrides audited?"],"valueHypothesis":"On-prem AI may speed triage if recommendations match officials; unquantified on this page.","potentialEngagement":"A shadow-mode agreement study on veterinary applications before expanding types.","unsupportedClaims":["Numeric processing-time reduction","Cross-license generalization","That open weights remove all vendor dependence"]},"engineering":{"fit":"Fits governments that must keep license packets on-prem without public-cloud LLM APIs.","architectureIntegration":"Qwen-VL extraction + Gemma 3 27B on-prem vLLM/H100 recommending issue/reject/manual review to officials.","prerequisites":["On-prem GPU capacity","Registry data access","Audit logging of recommendations"],"constraints":["Vendor case study evidence","Veterinary-only scope","No published baselines"],"security":"Enforce no-external-API controls and protect applicant PII in logs and training sets.","proofOfValue":["Shadow-mode agreement with officials","Time-to-first-feedback vs baseline","Override rate by reason"]},"delivery":{"work":"Stand up evaluation sets per license type and train officials on recommendation review.","dependencies":["Model ops / GPU platform","Registry stewards","License program owners"],"ownership":"Ministry digital leadership with program officials as decision owners.","skillsAdoption":"Officials need practice reviewing AI rationales and recording disagreements.","governanceCheckpoints":["Shadow-mode exit criteria","Type-expansion gate","Periodic audit-log review"],"acceptanceCriteria":["Proposed: no auto-issue until evaluation gates pass","Proposed: every override reason logged"],"risks":["Silent automation of denials","Drift on Ukrainian legal language","GPU capacity shortfalls"]}},"retrievedAt":"2026-11-06T22:06:48-06:00","enrichedAt":"2026-11-06T22:11:17-06:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Aimed at faster feedback to applicants and lower official workload; public accessibility of ePermit UX not discussed.","procurementImplications":"Open-model on-prem path rather than SaaS API; GPU datacenter capacity required; no commercial contract terms disclosed.","operatingModelImplications":"Officials still decide; AI triages completeness/consistency/eligibility; ministry developers fine-tune for Ukrainian language and document terms.","sourceVerification":{"openedUrl":"https://deepmind.google/models/gemma/gemmaverse/epermit/","referenceExcerpt":"cutting processing time dramatically and is running without latency issues","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}