Quantitative Developer – Agentic Research Lab بدوام كامل في دبي
التخصص
تكنولوجيا المعلومات
نوع الدوام
دوام كامل
الخبرة
أي مستوى
الموقع
دبي
وصف الوظيفة
About FlowdeskAt Flowdesk, we are building the infrastructure powering the next era of global finance. As a full-service digital asset trading and technology firm, we provide liquidity solutions, OTC trading, and market infrastructure that connect digital assets to global markets with institutional-grade precision, transparency, and scale. Flowdesk combines expertise from traditional finance, algorithmic trading, and blockchain technology to deliver always-on trading capabilities across the digital asset ecosystem.Founded in 2020, we have grown into a global team supporting institutions, exchanges, and token issuers worldwide as the financial system evolves toward tokenized markets.Why join Flowdesk? You’ll be joining a company operating at the intersection of finance, technology, and crypto - solving complex market problems, building critical infrastructure, and helping shape how digital assets move globally. We move fast, think long term, and work with a strong ownership mindset across engineering, trading, product, operations, and business teams.Quant Development is a new team inside Technology, owning the research platform that lets our quant researchers test thousands of hypotheses a day: the market-data store, the distributed compute systems, and the orchestration and tooling around them. Tick-level order-book data and our own execution traces stream through Redpanda into Delta Lake and ClickHouse, with a Kubernetes cluster running simulation campaigns on top.MissionYou’ll be the technical lead for the agentic research lab — setting the architecture, shipping the first production agents yourself, and defining the guardrails that keep an automated research loop honest. Agents formulate hypotheses, write and run experiments, check their own results, and watch live strategies for decay — humans review verdicts, not boilerplate. You’ll work closely with quant researchers, traders and the risk team, as well as the Systems and Research-tooling developers whose platform your agents run on.TasksMake the research platform production-grade (MLOps & CI/CD)Treat research like software: version, build, test and deploy every strategy, signal, dataset snapshot and agent prompt through CI/CD (GitHub Actions, GitOps with Flux / Argo CD)Own the model lifecycle end to end — experiment tracking, model registry, reproducible builds, automated validation gates, staged deployment and rollback — so a model that reaches capital always links back to its code, data and verdictPut the agents on the same rails: ship their tools, prompts and evals through CI, with a regression in an eval suite failing the PRPlug agents into the live data streamsConsume market data and execution traces from Kafka / Redpanda topics and query them at scale in ClickHouse, so monitoring and research agents see the same fresh data the desks see, within seconds rather than T+1Design the topics, schemas and materialized views the monitoring loop needs, with backpressure, replay and exactly-once semantics where they matterBuild the research agent loopDesign a multi-agent workflow: a hypothesis generator, an experiment writer that turns a hypothesis into a campaign spec, an executor that runs it on the simulation cluster, and a critic that tries to break the result before a human sees itExpose the platform to agents as typed tools — an MCP server over the research SDK, the campaign engine and the experiment registry — so agents call the same validated code paths researchers doMake “continuous” realRun scheduled research rounds (Dagster) on fresh data: re-validate live strategies, re-run parked ideas when the simulator improves, and propose parameter updatesLand every round in the experiment registry (MLflow-style) with its config hash, data snapshot and code version, so nothing an agent produces is unreproducibleBuild continuous monitoring with agentic triageTurn sim-vs-live comparisons, PnL, inventory, exposure, latency and regime indicators into alertsHave agents triage alerts first: gather context, check known causes, and draft an incident summary with a recommended action for the trader on callRun controlled experiments in productionOwn the path from shadow mode to paper trading to small live allocations, with explicit promotion criteria, kill-switches and rollback drills at every stageDesign the experimentation framework — A/B and interleaved parameter tests across venues and instruments — and the statistics that decide themGuard against automated overfittingBuild defences into the loop: a budget of trials per hypothesis, multiple-testing corrections (deflated Sharpe, false-discovery control), untouched holdouts agents cannot read, and independent re-execution of every claimed resultEvaluate the agents themselvesBuild eval suites with known-answer research tasks (planted signals, known bugs, look-ahead traps) and track how often agents find them, miss them or fake themRoute cheap models to extraction and classification, reserve frontier models for reasoning, and track cost per research roundRun models on our own hardwareStand up secure, on-prem GPU inference for open-weight models so proprietary data and strategy code never leave our infrastructureDefine which tasks may use external APIs and which must stay on-premWays of workingShip through code review and CI like the rest of engineering — no exceptions for agent-authored code or promptsDocument architecture decisions, promotion criteria and kill-switch conditions so another engineer can operate the loop without youOwn your systems end to end, from design through production monitoring and incident responseHow we’ll measure your impactResearch rounds completed per week without human authoring, and the share that end in a reviewed verdictShare of agent-reported results that reproduce on independent re-executionTime from a live-performance anomaly to a triaged summary in front of a traderStrategies promoted through the shadow → live path, and incidents caused by that path (target: zero)RequirementsMust Have10+ years building production software or ML systems, including several years as a technical lead who set architecture for other engineersAgents in production, not demos: multi-agent orchestration, tool calling, evaluation and red-teaming, cost and model routing, staged rollouts with a human in the loopHands-on quantitative experience: you have built backtests, factor or signal models, or risk models, and know how they lie — look-ahead bias, data snooping, regime change and transaction costs are second natureMLOps and CI/CD at platform level: you have built (not just used) model registries, experiment tracking, automated validation and deployment pipelines, and treat reproducibility as a hard requirementStreaming and analytical data systems in production: Kafka (or Redpanda) and ClickHouse, including schema design, partitioning, consumer lag and query performance at high volumeStrong Python and SQLAt ease on Linux: you debug on the box (processes, memory, networking, I/O, perf/strace-level tooling) and run workloads on Kubernetes with infrastructure-as-codeExperience shipping automated decision systems in regulated or capital-at-risk environments: kill-switches, audit trails, compliance reviewJudgement to know which parts of research to automate, and which must stay with a humanNice to HaveProduction experience in a systems language — Rust, C++ or Go — with Rust especially valuable, since the trading engine and simulator are written in itExperience in a trading environment, ideally market making, HFT or crypto (perps, funding, on-chain venues)Market microstructure knowledge: order books, queue position, markouts, adverse selectionMCP server design, or open-weight model serving (vLLM, SGLang or similar) on your own GPUsExperience with Delta Lake, Dagster, MLflow, GitOps (Flux / Argo CD) or TerraformBenefitsInternational environment (English is the main language)Team events and offsitesLearning and Development budgetHealth insurance for employee
التعليقات
لا توجد تعليقات منشورة حتى الآن.