Role Summary
We are looking for a Senior FullStack Engineer, AI to own and evolve Delta Exchange's suite of AI-powered products. You will work across multiple production AI applications: conversational support agents, RAG-powered search, code-generation copilots, and trading-strategy assistants. This is a hands-on role where you architect solutions around our existing stack, build retrieval pipelines, improve frontends with product sense, optimise inference costs, and run evaluations, collaborating closely with the product team.
Key Responsibilities
- Design, build, and maintain production AI applications end-to-end: backend, frontend, and inference services.
- Architect RAG systems using vector databases, embedding models, and chunking strategies optimised for accuracy and latency.
- Build agentic workflows with tool/function calling, multi-step reasoning, and structured output parsing, with accuracy and control as priority.
- Write and iterate on system prompts, few-shot examples, and prompt chains to maximise output quality.
- Implement function calling, tool-use patterns, and structured JSON/XML output handling using frontier and lightweight models from providers like Anthropic and OpenAI.
- Drive cost optimisation: model selection, caching, token budgeting, and request batching at scale.
- Build and maintain evaluation frameworks to measure accuracy, relevance, hallucination rates, and regression across prompt and model changes. Experience with observability tools (Sentry, Opik, etc.) is a must.
- Work with message queues (RabbitMQ), caching layers (Redis), and relational databases (PostgreSQL) powering AI service backends.
- Deploy and manage AI services on Kubernetes with CI/CD pipelines on AWS/GCP.
- Integrate AI capabilities with third-party platforms (Telegram bots, chat widgets, etc.).
- Contribute to architectural decisions: model selection, hosting (cloud APIs vs. self-hosted), and build-vs-buy trade-offs.
Required Skills & Experience
- 5+ years shipping production software systems.
- 2 years building AI/LLM-powered applications end-to-end with real users and volume. Not prototypes.
- Strong experience with RAG architectures: vector databases, embedding models, chunking/indexing strategies, and retrieval evaluation.
- Deep understanding of LLM capabilities and limitations: prompt engineering, function/tool calling, structured outputs, context window management, and multi-turn conversations.
- Experience with LLM provider APIs and abstraction layers (OpenAI, Anthropic, LiteLLM, OpenRouter, or similar).
- Proficiency in Python (Flask/FastAPI) and/or Node.js/TypeScript (Next.js, Vercel AI SDK). Golang experience is a plus.
- Hands-on experience building evals, tracking quality metrics, and debugging non-deterministic outputs in production.
- Familiarity with cost optimisation: model routing, caching, token usage monitoring, and prompt compression.
- Solid fundamentals in data structures, algorithms, and system design.
- Experience with containerised deployments (Docker, Kubernetes) and cloud platforms (AWS/GCP). Practical understanding of k8s concepts and trade-offs is a must.
Nice to Have
- Experience with agentic frameworks (Vercel AI SDK, LangChain, Mastra, etc.). Even better if you have built Gen AI apps using raw HTTP calls to provider APIs and designed efficient conversation persistence to a database.
- Background in fine-tuning or training open-source models. Huge plus if you can demonstrate matching proprietary model quality (Sonnet 4.6, Haiku 4.5, etc.) with fine-tuned alternatives.
- Knowledge of cryptocurrency, derivatives trading, or financial systems. Helps understand how AI can improve product UX.
- Open-source contributions or personal projects with real traction. Even a tool you built to solve your own problem counts.
- Your day is spent mostly in AI coding harnesses (Claude Code, Codex, Droid, etc.), MCP servers, custom skills, and similar tooling. We use AI to build AI, and candidates already living this workflow will ramp up fast.
- Proven ability to debug production AI systems: diagnosing tool call failures, optimising function calling patterns, refining prompts and tool descriptions under real traffic.
- Extreme ownership and bias towards action. You treat production AI systems as your own, proactively improving quality, latency, and cost without waiting to be asked. You deliver your best work especially when no one is watching.