AI Agentic integration Engineer (Senior / Lead)

Vallettasoftware - custom mobile/web software developer in the US and Europe. Our teams implement IT projects of varying complexity, including website and mobile app development, enterprise systems, and solutions based on artificial intelligence and machine learning (AI/ML).

Our company has earned a place among Clutch's Fall 2025 Champions. These awards confirm that we are a Top 15 AI Agent Developing companies!

We are a distributed team - you can work from any place in the world, except RF and RB, ensuring silence, good internet connection, availability and proper environment .

We are looking for a Senior / Lead AI Engineer to build production-ready AI systems where:

LLM is the core layer of the solution

Agentic workflows are used as the primary orchestration pattern

System quality is managed through evaluation

Reliability, observability, and cost control are designed as part of the architecture, not added later

This is not a backend engineer with AI functions, nor is it a prompt engineer.

This is an engineer who knows how to build AI-native systems end-to-end and take them all the way to production.

1. Hard Requirements

1.1. AI / LLM Systems

Must-have:

Real experience developing and shipping production LLM systems

Experience working with LLM APIs: OpenAI / Anthropic / Gemini / similar

Prompt design

Structured outputs

Tool / function calling

Model selection and understanding of trade-offs between models

1.2. Agentic Systems

Must-have:

Experience designing multi-step workflows

Experience developing agent-based systems: single-agent and/or multi-agent

Orchestration: planning, execution, retry, fallback, verification

Management of:

State

Context

Memory

Understanding when an agentic approach is needed and when it's not

Understanding of trust boundaries in agentic systems

Principle of least privilege for tool permissions

Protection against indirect prompt injection via external data (retrieval, tool results, external APIs)

1.3. Evaluation & Quality Control

Must-have:

Building evaluation pipelines

Offline evaluation

Comparing prompt / model versions

Quality metrics

Approaches to online validation / A/B testing / human review loops

Ability to connect evaluation to real product quality

1.4. Context Management & Hallucination Control

Must-have:

Context management:

Chunking strategies

Context window optimization

Memory patterns

Retrieval scope control

Hallucination reduction techniques:

Grounding

Retrieval

Tool-based verification

Constraints

Self-check / validation patterns

1.5. Production / LLM Ops / Reliability

Must-have:

Retries / exponential backoff

Timeout handling

Fallbacks / model routing

Degraded mode / graceful failure

Rate-limit handling

Observability:

Latency

Token usage

Cost

Failure rate

Output quality signals

Cost control

Monitoring and debugging AI systems in production

PII handling: filtering before logging, tenant isolation in memory and retrieval

Output validation and content guardrails

Awareness of data residency risks when using external LLM APIs

1.6. Data / Retrieval

Must-have:

Understanding of retrieval pipelines:

Embeddings

Chunking

Reranking

Retrieval quality tuning

Experience with vector storage / vector DB of any type

Working with structured and unstructured data

1.7. Engineering Foundation

Must-have:

Strong engineering background in backend / system development

Backend stack is not critical

Ability to build APIs, services, integrations, async workflows

SQL + NoSQL

Git, Docker, CI/CD

Nice-to-have:

Full-stack development experience (backend + frontend)

Understanding of UI/UX aspects of AI products (chat, copilots, dashboards)

1.8. AI Safety & Security

Must-have (awareness level for Senior, ownership for Lead):

Prompt Injection:

Understanding the difference between direct and indirect injection

Indirect injection — a specific threat to agentic systems: attacks via data from retrieval, tool results, external sources

Content sanitization before inserting into context

Architectural separation of system prompt, user input, and external data

Tool Permission Model:

Principle of least privilege: minimum necessary permissions for each agent and tool

Separation of read-only and write operations at the architecture level, not via prompts

Human-in-the-loop for irreversible actions (delete, send, publish, execute code)

Whitelist of allowed external calls

Data Leakage & PII:

Tenant isolation: one user's data never enters another user's context

PII masking before logging (including tool results and retrieval results)

Understanding that LLM APIs are third-party; for regulated domains — DPA, filtering, or self-hosted

Output Safety:

Output validation: schema checks, content filtering

Understanding the difference between "the prompt says don't do X" (weak protection) and "the architecture does not allow X" (strong protection)

1.9. Communication

Must-have:

English B2+

Ability to explain architectural decisions, trade-offs, and risks

For Lead: ability to set engineering standards and lead the technical direction of the team