Gigi is the agentic operating system of enterprise media buying. Our first product is the AI media manager for Amazon DSP. Since launching it in summer 2025, we've grown to manage hundreds of millions in advertising spend on behalf of the largest agencies in the world.
We're backed by top venture investors Golden Ventures and Aperiam Ventures. Our founding team has spent 15+ years building in Amazon Ads and ecommerce, including a prior $100M+ exit.
We're hiring an AI Agent Engineer to join our engineering team in Toronto and own end-to-end development of the agents and workflows behind Gigi.
This is fundamentally a software engineering role. Much of the work will feel familiar: designing APIs, modelling data, integrating services, writing tests, debugging production issues, and making architectural decisions. Your focus will be how LLMs reason, use tools, retain context, and take actions on behalf of customers—and the software that makes those actions reliable.
You do not need to be an AI researcher or have years of professional LLM experience. Strong software engineers who have started building with LLMs and agents, and want to take that work further, should apply. You'll use tools such as Codex, Claude Code, and Cursor as a core part of how you work, while building across agent logic, backend services, data, and the customer-facing product.
As an early employee, you will have a meaningful role in shaping our product, engineering culture, and company.
WHAT YOU'LL DO
- Evals. Build a sophisticated eval framework across every Gigi use case, quantitative and qualitative, tied to real outcomes (spend, ROAS, margin, profit). Deterministic checks, LLM-as-judge graders calibrated against human review, trajectory-level evals for multi-step runs, and replay against historical account data. No model, prompt, or harness change ships without passing them.
- Context engineering. What goes into the context window on every call: advertiser context, account state, AMC results, retail signals, agency best practices, and prior decisions. You'll own retrieval, compaction for long-running trajectories, and cache-aware prompt design.
- Memory. Short-term, long-term, and cross-session memory for every brand, agency, and media buyer. The hard part is deciding what is worth remembering, so you'll build the agents that distill approvals, rejections, and outcomes into durable memory and handle conflicts and staleness. Memory stays model-agnostic and strictly separated per customer, including customers that require no training on their data.
- Agent behavior. Sub-agents, how tools are presented to the agent, and the line between what the LLM reasons about and what deterministic code executes. Risk and confidence scoring, guardrails, and how Gigi earns more autonomy with each customer. When a new model makes a workaround unnecessary, you delete it.
- Continual learning and the moat. The loop around the agent: capture traces, extract learnings, update context, memory, harness, or model weights, measure, repeat. Every action should leave Gigi smarter and more personalized for that customer, and outcomes across all customers should make every recommendation better.
- Model routing. Pick the right model for every call on price, performance, and speed, inside the harness, with fallbacks across providers. Stay model-agnostic, evaluate new models within hours of release, and own model spend.
- Post-training. The strategy and the implementation. Start with narrow, high-volume tasks (classification, extraction, summarization, guardrails, routing) and move toward the core agent loop when the data supports it, using fine-tuning, preference optimization, distillation, and RL against outcome-based rewards. You'll build the training data from traces and make the build vs buy calls on training and serving with engineering leadership.
WHAT SUCCESS LOOKS LIKE
- First 30 days. You've shipped your first improvements to production, stood up trusted evals on our core…