VentraGate
Website:
ventragate.com
Job details:
Company Description
VentraGate is a cloud-focused solutions and services firm dedicated to driving digital transformation through modern, cloud-based technologies that customers find intuitive and valuable. The company specializes in imagining, building, and sustaining scalable digital solutions that help organizations modernize their operations. As a leading transformation consultancy, VentraGate partners with enterprises of all sizes to become faster, more agile, and more competitive in their markets. Team members at VentraGate work in a collaborative environment that emphasizes innovation, technical excellence, and long-term client success.
Role Description : "This is a product engineering role, not a research role. You will spend your time turning models into reliable systems: retrieval pipelines, agentic workflows, tool integrations, evaluation harnesses, and the guardrails that keep all of it trustworthy under real traffic.
You'll own features end to end, from prototype to production. And you'll be the person who knows whether the system actually works because you built the evals that prove it".
What you will do :
- Design, build, and ship LLM-backed product features: RAG systems, agentic workflows, tool and function calling, structured extraction, and multi-step reasoning pipelines.
- Build evaluation harnesses offline eval sets, LLM-as-judge rubrics, regression suites, online A/B tests so prompt and model changes ship on evidence rather than intuition.
- Own context engineering: chunking strategy, embedding and reranking choices, hybrid search, caching, and context-window budgeting.
- Integrate models into production services with sane latency, cost, and failure behavior streaming, retries, timeouts, cross-provider fallback, and token and cost instrumentation.
- Instrument and monitor AI in production: tracing, prompt and response logging, quality and drift dashboards, and incident response when output quality regresses.
- Implement safety and reliability guardrails input and output validation, PII handling, prompt-injection defenses, jailbreak resistance, and human-in-the-loop escalation.
- Fine-tune or adapt models (LoRA, SFT, distillation to smaller models) where it beats prompting on cost or quality.
- Work with product, design, and domain experts to define what “good output” means before the prompt gets written.
- Write clear technical docs and post-mortems, and help teammates new to LLM systems get productive.
What we're looking for :
- 2–5 years of professional software engineering experience, including at least one year shipping LLM features to production users.
- Strong Python, Solid fundamentals: data structures, concurrency, API design, testing.
- Hands-on experience with modern LLM APIs and SDKs Anthropic Claude, OpenAI, Gemini, or open-weight models served through vLLM, Ollama, Bedrock, or Vertex.
- Practical command of prompt and context engineering: system prompts, few-shot design, structured and JSON output, tool use, and reflection or self-critique patterns.
- Real RAG experience: embeddings, vector stores (pgvector, Pinecone, Weaviate, Qdrant, Milvus), chunking strategy, hybrid and BM25 search, and reranking.
- Demonstrated ability to evaluate AI systems. You can describe an eval set you built, the metrics you chose, and a change you rejected because the numbers said so.
- Experience with agent orchestration LangGraph, LlamaIndex, the Vercel AI SDK, DSPy, the Claude Agent SDK, or a hand-rolled equivalent and with MCP or a comparable tool-integration standard.
- Production engineering skills: Docker, CI/CD, a major cloud, observability, and cost awareness at scale.
- Fluency in AI-specific failure modes hallucination, prompt injection, data leakage, non-determinism and concrete mitigations for each.
- Clear written communication. You can explain a tradeoff to a PM and a stack trace to an engineer.
Nice to have
- Fine-tuning experience (LoRA/QLoRA, SFT, DPO or RLHF) with PyTorch or Hugging Face.
- Inference optimization: quantization, batching, KV caching, speculative decoding, GPU serving with vLLM, TGI, or TensorRT-LLM.
- LLM observability tooling LangSmith, Langfuse, Braintrust, Arize, or W&B Weave.
- Multimodal work: vision, audio, document understanding, or OCR pipelines.
- Voice and realtime AI, streaming architectures, or other low-latency interactive systems.
- Classical ML background feature engineering, gradient boosting, ranking or recommender systems.
- Data engineering: dbt, Airflow or Dagster, Spark, warehouse modeling.
- Working knowledge of AI governance and regulation (EU AI Act, NIST AI RMF, SOC 2, GDPR) and how it shapes system design.
Click on Apply to know more.
