Teknologji Informacioni (IT)Prishtinë3 ditë më parëFull Time
Applied AI Engineer
Omzo is seeking a Staff Applied AI Engineer to create the AI backend for Omzo Air, taking it from managed-model assessment to a safe, measurable production rollout for a health platform subject to HIPAA requirements.
About Omzo
Omzo is developing Omzo Air, an AI assistant designed to help users interpret approved health and wearable data, identify suitable next actions, and reach a human care team when needed.
Omzo Air is not designed to function as an autonomous clinician. The product is being built for production use from day one, emphasizing measurable safety, dependable operation, managed costs, and well-defined clinical limits.
About the role
The company expects to use a managed model, as this is the most practical approach for satisfying HIPAA, security, and reliability needs without running its own LLM infrastructure.
OpenAI API and Azure OpenAI are currently the main options, although the deployment decision is still open. You will direct the technical assessment and advise on the provider, model, and configuration according to quality, safety, latency, regional availability, cost, and operational suitability.
Privacy, Security, and Legal teams will sign off on contractual and data-handling obligations. Clinical leadership will authorize intended use, clinical-safety policies, and release requirements.
You will create and take responsibility for the AI backend service. Omzo’s backend, mobile, and web engineers will consume the internal API you deliver while retaining ownership of the related application and interface code. Platform/SRE is responsible for cloud infrastructure, networking, deployments, and infrastructure incidents.
Healthcare background is beneficial but not mandatory. Experience collaborating with domain specialists on regulated, safety-critical, or sensitive-data systems is important. You will not be required to decide clinical policy on your own.
What you will build and own
Direct the technical comparison of OpenAI API, Azure OpenAI, and any authorized alternatives.
Develop the AI backend service and a dependable internal API for Omzo’s backend, mobile, and web teams.
Create the model gateway, covering routing, streaming, timeouts, limited retries, rate limits, authorized caching, cost measurement, and safe degraded operation.
Take ownership of model-related response performance while partnering with Platform/SRE on infrastructure capacity, networking, and provider quotas.
Implement retrieval from clinician-approved materials, with ingestion, provenance, review state, versioning, citations, and rollback support.
Implement task and risk routing that can:
Select a smaller, faster model for validated work such as intent classification, retrieval preparation, and approved FAQs.
Teknologji Informacioni (IT)Prishtinë3 ditë më parëFull Time
Applied AI Engineer
Omzo is seeking a Staff Applied AI Engineer to create the AI backend for Omzo Air, taking it from managed-model assessment to a safe, measurable production rollout for a health platform subject to HIPAA requirements.
About Omzo
Omzo is developing Omzo Air, an AI assistant designed to help users interpret approved health and wearable data, identify suitable next actions, and reach a human care team when needed.
Omzo Air is not designed to function as an autonomous clinician. The product is being built for production use from day one, emphasizing measurable safety, dependable operation, managed costs, and well-defined clinical limits.
About the role
The company expects to use a managed model, as this is the most practical approach for satisfying HIPAA, security, and reliability needs without running its own LLM infrastructure.
OpenAI API and Azure OpenAI are currently the main options, although the deployment decision is still open. You will direct the technical assessment and advise on the provider, model, and configuration according to quality, safety, latency, regional availability, cost, and operational suitability.
Privacy, Security, and Legal teams will sign off on contractual and data-handling obligations. Clinical leadership will authorize intended use, clinical-safety policies, and release requirements.
You will create and take responsibility for the AI backend service. Omzo’s backend, mobile, and web engineers will consume the internal API you deliver while retaining ownership of the related application and interface code. Platform/SRE is responsible for cloud infrastructure, networking, deployments, and infrastructure incidents.
Healthcare background is beneficial but not mandatory. Experience collaborating with domain specialists on regulated, safety-critical, or sensitive-data systems is important. You will not be required to decide clinical policy on your own.
What you will build and own
Direct the technical comparison of OpenAI API, Azure OpenAI, and any authorized alternatives.
Develop the AI backend service and a dependable internal API for Omzo’s backend, mobile, and web teams.
Create the model gateway, covering routing, streaming, timeouts, limited retries, rate limits, authorized caching, cost measurement, and safe degraded operation.
Take ownership of model-related response performance while partnering with Platform/SRE on infrastructure capacity, networking, and provider quotas.
Implement retrieval from clinician-approved materials, with ingestion, provenance, review state, versioning, citations, and rollback support.
Implement task and risk routing that can:
Select a smaller, faster model for validated work such as intent classification, retrieval preparation, and approved FAQs.
Select a more capable model for complex work that remains inside the authorized product scope.
Decline or constrain requests outside the approved scope.
Route high-risk, uncertain, or clinician-only questions to a human care team.
Build cost efficiency into every layer and use models as a last option rather than the default:
Process predictable requests through regular application code rather than model calls, including greetings, menu-based navigation, appointment and shipping status, account queries, and answers that can be directly retrieved from internal systems.
Provide approved FAQs and recurring questions through a reviewed answer library or semantic cache, using a model only when an approved match is unavailable.
Apply deterministic rules and lightweight classifiers ahead of model calls so straightforward intents do not reach a paid endpoint.
Limit prompts and context through cached system prompts, bounded conversation history, only the retrieved sources required for an answer, and capped output size.
Measure cost per successful task for each route, then routinely shift workloads from models to code, from larger to smaller models, or from live calls to batch processing when evaluations show no reduction in quality or safety.
Create multi-layered safety measures based on clinician-approved rules, classifiers, and conversation context.
Develop response-release safeguards. Responses involving policy sensitivity or elevated risk must be reviewed before release, using buffered or validated segment streaming when suitable.
Build the human-escalation process, including ticket generation, priority assignment, minimized context, delivery verification, retries, status monitoring, and an approved temporary response for the patient.
Create clinician-reviewed evaluation suites for retrieval, grounding, citations, risk identification, routing, refusal behavior, supported languages, prompt injection, and regressions.
Conduct evaluations for every applicable change and on a regular schedule with synthetic, de-identified, or otherwise authorized cases.
Implement privacy-safe monitoring of model quality, latency, reliability, escalation behavior, and cost without placing protected health information in ordinary logs or traces.
Lead model and prompt version control, shadow evaluations, canary releases, and rollback together with Platform/SRE.
Assess production quality weekly alongside the clinical lead and convert findings into measurable engineering enhancements.
Set standards for the AI codebase and mentor engineers who later join the AI effort.
A more powerful model cannot replace a doctor. Routing must assess clinical risk independently of technical difficulty, and each route must clear its own safety and quality assessments. Clinical-safety thresholds cannot be compromised for cost reductions.
Fallback traffic can only go to model deployments that have separately met contractual, regional, security, and clinical-evaluation standards. Where no approved fallback exists, the system must provide a safe degraded response along with the suitable human or emergency route.
What we are looking for
Proven Staff-level technical leadership and substantial hands-on responsibility for at least one production LLM system serving actual customers, generally backed by six or more years of relevant engineering work.
Production experience with structured outputs, retrieval-augmented generation, model evaluation, and managed model APIs.
Background building internal APIs or backend services consumed by other application teams.
Experience architecting reliable services using databases, queues, rate limits, timeouts, retries, and external API dependencies.
A demonstrated history of applying evaluations to guide model, prompt, and routing choices.
Experience lowering production AI costs through routing, caching, prompt design, or replacing model requests with code while preserving quality.
Experience safeguarding sensitive data in a regulated or security-aware setting.
Capacity to lead technical design and communicate trade-offs clearly to engineering, product, clinical, and leadership stakeholders.
How we work
Releases are delivered in stages with defined entry and exit conditions.
Before release, every model, prompt, route, retrieval, policy, or knowledge modification must pass the relevant evaluation gates.
Measured evidence is favored over provider benchmarks and intuition.
Clinician-reviewed scoring rubrics are used instead of presuming each health question has a single precise answer.
Production quality is assessed weekly with the clinical lead.
Significant provider, architecture, and clinical-safety decisions are documented.
Select a more capable model for complex work that remains inside the authorized product scope.
Decline or constrain requests outside the approved scope.
Route high-risk, uncertain, or clinician-only questions to a human care team.
Build cost efficiency into every layer and use models as a last option rather than the default:
Process predictable requests through regular application code rather than model calls, including greetings, menu-based navigation, appointment and shipping status, account queries, and answers that can be directly retrieved from internal systems.
Provide approved FAQs and recurring questions through a reviewed answer library or semantic cache, using a model only when an approved match is unavailable.
Apply deterministic rules and lightweight classifiers ahead of model calls so straightforward intents do not reach a paid endpoint.
Limit prompts and context through cached system prompts, bounded conversation history, only the retrieved sources required for an answer, and capped output size.
Measure cost per successful task for each route, then routinely shift workloads from models to code, from larger to smaller models, or from live calls to batch processing when evaluations show no reduction in quality or safety.
Create multi-layered safety measures based on clinician-approved rules, classifiers, and conversation context.
Develop response-release safeguards. Responses involving policy sensitivity or elevated risk must be reviewed before release, using buffered or validated segment streaming when suitable.
Build the human-escalation process, including ticket generation, priority assignment, minimized context, delivery verification, retries, status monitoring, and an approved temporary response for the patient.
Create clinician-reviewed evaluation suites for retrieval, grounding, citations, risk identification, routing, refusal behavior, supported languages, prompt injection, and regressions.
Conduct evaluations for every applicable change and on a regular schedule with synthetic, de-identified, or otherwise authorized cases.
Implement privacy-safe monitoring of model quality, latency, reliability, escalation behavior, and cost without placing protected health information in ordinary logs or traces.
Lead model and prompt version control, shadow evaluations, canary releases, and rollback together with Platform/SRE.
Assess production quality weekly alongside the clinical lead and convert findings into measurable engineering enhancements.
Set standards for the AI codebase and mentor engineers who later join the AI effort.
A more powerful model cannot replace a doctor. Routing must assess clinical risk independently of technical difficulty, and each route must clear its own safety and quality assessments. Clinical-safety thresholds cannot be compromised for cost reductions.
Fallback traffic can only go to model deployments that have separately met contractual, regional, security, and clinical-evaluation standards. Where no approved fallback exists, the system must provide a safe degraded response along with the suitable human or emergency route.
What we are looking for
Proven Staff-level technical leadership and substantial hands-on responsibility for at least one production LLM system serving actual customers, generally backed by six or more years of relevant engineering work.
Production experience with structured outputs, retrieval-augmented generation, model evaluation, and managed model APIs.
Background building internal APIs or backend services consumed by other application teams.
Experience architecting reliable services using databases, queues, rate limits, timeouts, retries, and external API dependencies.
A demonstrated history of applying evaluations to guide model, prompt, and routing choices.
Experience lowering production AI costs through routing, caching, prompt design, or replacing model requests with code while preserving quality.
