[2027 Summer] AI Research and Engineering Intern for Business Intelligence (8023)

Location: San Jose, CA, US

[2027 Summer] AI Research and Engineering Intern for Business Intelligence (8023)

Overview of Role

Frontier models can now reason, write code, and use tools. The next step is getting agents to work reliably over long tasks with real data, and serving them efficiently in production. Enterprise business intelligence is one of the hardest places to test this. Agents must work through messy schemas and ambiguous questions. They have to keep business logic consistent across many steps, and they have to give answers someone can check. At a company operating at TSMC's scale, wrong numbers have real costs.

AI4BI Innovation center is looking for exceptional PhD and Master's students to work on this problem with us. You will own a research and engineering question end-to-end. That means designing the method, building the agent harness and environment, training or adapting models, deploying them, and measuring progress with evaluations that hold up. The goal is a first-author paper at a top venue (e.g., NeurIPS, ICLR, ICML, ACL, EMNLP) or a prototype that ships into real decision-making workflows for business stakeholders. Many projects do both.

This is not a typical software engineering internship. You will have dedicated mentorship from AI Scientists and Data engineers, and access to realistic enterprise-scale data and business problems that are hard to get in academia.

What You'll Work On

Your project will fall within one or more of these directions. Each is grounded in real business problems.

Agent harnesses and long-horizon agents. Design the harness around the model: the agent loop, tool and sandbox interfaces, context management and compaction, memory, sub-agent delegation, permissions, and error recovery. Build agents that can investigate a business question across hours of work. They explore data, form hypotheses, run analyses, and produce a defensible report.

Tool use and multi-agent systems. Design agents that operate over warehouses, semantic layers, dashboards, and internal APIs through standard tool protocols such as the Model Context Protocol (MCP). Study when multi-agent orchestration helps a single strong agent, and when it gets in the way.

Pretrained time-series foundation models. Evaluate, adapt, and fine-tune pretrained time-series foundation models (e.g., TimesFM, Chronos, Moirai) for forecasting demand, capacity, and supply-chain signals. Topics include zero-shot versus fine-tuned performance, covariates and multivariate forecasting, probabilistic forecasts and calibration, and domain-adaptive pretraining on enterprise time series. Study how LLM agents can call, combine, and explain these forecasting models, alongside optimization and network science methods.

AI infrastructure and deployment. Build the systems that take models and agents from prototype to production. Topics include efficient LLM serving and inference optimization (batching, KV-cache management, quantization, speculative decoding), model routing and cost/latency trade-offs, on-premise and secure deployment, and observability and tracing for agent runs (including guardrails and monitoring).

Agent environments and evaluation science. Build realistic, reproducible environments and benchmarks for data agents. Develop reliable LLM-as-judge and rubric-based evaluators. Measure what matters: correctness, consistency, calibration, latency, cost, and trustworthiness.

Synthetic data and post-training. Generate verifiable training and evaluation tasks at scale. Adapt models to the enterprise domain through fine-tuning, distillation into small, cost-efficient models that can be deployed, and, where useful, reinforcement learning from verifiable signals such as SQL results and metric consistency checks.

Responsibilities

Lead a research and engineering project. Frame an open problem, form hypotheses, and run careful experiments, from small-scale ablations up to training and serving on GPUs.

Develop models and agents. Build and manage advanced machine learning, deep learning, LLM, foundation model, and time-series methods. Use them to develop data-driven solutions to complex business problems, and implement agent harnesses, tool interfaces, and environments grounded in realistic BI workloads.

Deploy and operate. Package, serve, and monitor models and agents in production-like environments. Optimize them for reliability, latency, and cost.

Turn data into decisions. Analyze and interpret complex data sets. Translate the results into actionable insights that improve decision-making for business stakeholders.

Evaluate rigorously. Design benchmarks and evaluation protocols that measure reliability, not just demo quality. Report negative results honestly.

Collaborate across teams and borders. Work with cross-functional and international teams in the US and Taiwan to find and prioritize opportunities to use data. Partner with research scientists and data engineers to develop and deploy predictive models and agentic systems. Help build and maintain efficient processes for collecting, storing, and analyzing data. Take part in weekly research syncs, paper reading groups, and brainstorming sessions.

Communicate your work. Present insights and recommendations to both technical and non-technical stakeholders through internal reports and presentations. Write up findings for publication and/or internal deployment.

Minimum Qualifications

Education: Currently pursuing a Ph.D. or M.S. in Computer Science, Electrical Engineering, Information Systems, Information Science, Statistics, or a related AI/ML field.

AI-enabled analytics knowledge: Depth in one or more of machine learning, deep learning, large language models, AI agents, time-series modeling and forecasting, optimization, network science, or ML systems.

Engineering: Strong Python skills and hands-on experience building and shipping ML or LLM systems.

Research ability: Able to read recent papers critically, then implement and extend th