Understanding the physics of the early universe remains one of the central challenges in modern cosmology. With a new generation of increasingly precise CMB surveys on the horizon, we are approaching the sensitivity required to address many of these open questions. To fully capitalise on this opportunity, we need robust and efficient pipelines that can reliably extract early-universe signatures from high-fidelity CMB data. In this talk, I will present three complementary efforts toward this goal. First, I will describe progress in mapmaking pipeline development using FURAX, an open-source, GPU-optimized framework built on JAX. Its modular architecture enables the seamless incorporation of instrumental systematics, including non-ideal half-wave plate behavior and asymmetric beams. Second, I will discuss studies in map-based component separation and tensor-to-scalar ratio estimation. Using the spectral-likelihood formalism, we account for spatially varying foreground spectra and incorporate the full observation matrix for robust estimation of r. Finally, I will present a general framework for estimating arbitrary shapes of primordial non-Gaussianity using CMB-BEST, which we use to derive comprehensive constraints on cosmological collider models, yielding the strongest bounds to date. I will conclude with future prospects and the key challenges that must be addressed to unlock the full potential of future CMB observations.
