The trigger and data acquisition system is an essential and evolving component of the ATLAS detector. The sequential hardware and software triggers are responsible for the real-time reconstruction of all physics signatures of interest and the selection of events based on diverse trigger hypotheses. Operational constraints enforce stringent requirements on the data throughput and latency of trigger reconstruction and impose strict kinematic thresholds on physics objects.This work centres on an unconventional analysis strategy, Trigger-Level Analysis (TLA) that avoids the trigger-imposed statistical limitations of conventional resonance searches. TLA extends the reach of hadronically decaying dark matter candidates towards the electroweak scale and exploits the luminosity provided by the LHC by utilising a partial event read-out system. In the TLA paradigm only the candidate objects reconstructed in the trigger are recorded, reducing the memory footprint of each event and permitting an increased event rate to be stored for subsequent analysis. The TLA presented in this thesis performs a heavy-flavoured dijet resonance search using proton-proton collisions at a centre-of-mass energy of $\sqrt{s} = 13.6$TeV corresponding to an integrated luminosity of $25$\,$fb^{-1}$. This work includes the first usage of a fast, trigger-optimised jet flavour tagging algorithm to identify the beauty quark-initiated jet resonance. This search is the first ATLAS analysis to use symbolic regression to enhance its data-driven background estimation strategy. Improvements in the background modelling with respect to the standard functional method are demonstrated. At the partial unblinding stage no significant excess over the background estimate is observed and expected exclusion limits on a vector-axial dark matter mediator coupling to beauty quarks are derived in the mass range 125-400 GeV. This TLA produces competitive exclusion limits with other collider searches whilst using a dataset 5 times smaller; with a dataset of comparable size TLA is projected to set the strongest ATLAS limits using a resolved dijet system in the low mass regime. Looking ahead to the High Luminosity upgrade to the LHC, the ATLAS trigger and data acquisition system will undergo a comprehensive upgrade. In order to fully exploit the unprecedented luminosity delivered to the ATLAS experiment, the real-time reconstruction must maintain its high signal efficiency in the presence of an increased background rate. Fast Machine Learning can offer potential alternatives to existing iterative trigger preselection reconstruction algorithms, particularly using low-level detector information as input. This thesis introduces the design and implementation of an object detection CNN for hadronic jet finding in the ATLAS calorimeter. The custom architecture, CaloJetSSD, is used to identify and localise jets and to subsequently estimate their transverse momenta in real-time. This work includes an evaluation of the performance of the CaloJetSSD architecture on a set of simulated particle interactions in the ATLAS detector with up to 200 concurrent pile-up interactions. CaloJetSSD achieves comparable trigger performance to the current ATLAS jet trigger reconstruction algorithms while achieving a speed-up of almost a factor 10. These studies point to significant reductions in the execution time of calorimeter-only jet trigger preselections and a more pile-up robust hadronic trigger reconstruction.