SIGNAL

Self-improving AI system for paid media, using Bayesian optimisation, LLM agents, and computer vision to drive autonomous growth decisions.

Comment

Inspiration

Marketing optimisation is still largely manual, reactive, and fragmented across platforms. We wanted to explore whether a multi-agent AI system could continuously analyse signals, learn from outcomes, and autonomously optimise campaign performance.

What it does

SIGNAL is an agentic ad optimisation platform that combines specialised AI agents, probabilistic optimisation, and real-time market intelligence to improve paid media performance.

Features:

  • 5 AI agents (Bidding, Creative, Content, Market Intelligence, Orchestration)
  • Autonomous agent learning loops
  • DisBERT commercial intent classification
  • LOESS bid curve modelling
  • James-Stein Bayesian cold-start optimisation
  • Tavily-powered market research
  • LLM-as-a-Judge evaluation framework
  • Full observability and tracing

How I built it

Stack: FastAPI, React, PostgreSQL, SciPy, DisBERT, Overmind SDK, Tavily, OpenAI, Anthropic.

Pipeline: Research → Classification → Bid Discovery → Bayesian Optimisation → Agent Evaluation → Agent Learning → Autonomous Execution

Agents are orchestrated using Overmind SDK, automatically tuned through Overmind Agent Optimizer, and continuously improved via closed-loop evaluation feedback.

Challenges I ran into

  • Coordinating multiple agents with conflicting recommendations
  • Building reliable evaluation and feedback loops
  • Optimising campaigns with sparse cold-start data
  • Balancing autonomous behaviour with decision quality and control

Accomplishments that I’m proud of

🏆 1st Place — AdTech Cursor Hackathon London

  • Built a production-style multi-agent optimisation architecture
  • Implemented autonomous agent learning loops
  • Combined probabilistic modelling, LLM orchestration, and optimisation algorithms into a unified decision system

What I learned

The biggest insight was that agent performance depends less on the model itself and more on orchestration, evaluation, observability, and feedback loops. Classical optimisation techniques remain highly effective when combined with modern AI systems.

What’s next for SIGNAL

  • Direct Google Ads and Meta integrations
  • Reinforcement learning for budget allocation
  • Computer vision for creative analysis and fatigue detection
  • Long-term agent memory and learning
  • Fully autonomous campaign execution

Vision: Build the operating system for autonomous marketing optimisation.

Built With

Share this project:

Updates