While a featureless, nearly scale invariant, primordial scalar power spectrum fits the most recent Cosmic Microwave Background (CMB) data rather well, certain features in the spectrum are known to lead to a better fit to the data (although, the statistical significance of such results remains an open issue). In the inflationary scenario, one or more periods of deviations from slow roll are necessary in order to generate features in the scalar perturbation spectrum. Over the last couple of years, it has been recognized that such deviations from slow roll inflation can also result in reasonably large non-Gaussianities. The Starobinsky model involves the canonical scalar field and consists of a linear inflaton potential with a sudden change in the slope. The change in the slope causes a brief period of departure from slow roll which, in turn, results in a sharp rise in power, along with a burst of oscillations in the scalar spectrum for modes that leave the Hubble radius just before and during the period of fast roll. The hallmark of the Starobinsky model is that it allows the scalar power spectrum to be evaluated analytically in terms of the three parameters that describe the model, viz. the two slopes that describe the potential on either side of the discontinuity and the Hubble scale at the time when the field crosses the discontinuity. In this work, we evaluate the bi-spectrum of the scalar perturbations in the Starobinsky model in the equilateral limit. Remarkably, we find that, just as the power spectrum, all the different contributions to the the bi-spectrum too can be evaluated completely analytically and expressed in terms of the three paramaters that describe the model. We show that the quantity fNL, which characterizes the extent of non-Gaussianity, can be expressed purely in terms of the ratio of the two slopes on either side of the discontinuity in the potential. Further, we find that, for certain values of the parameters, fNL in the Starobinsky model can be as large as the mean value that has been arrived at from the analysis of the recent CMB data. We also demonstrate that the usual hierarchy of contributions to the bi-spectrum can be altered for certain values of the parameters. Altogether, we find that the Starobinsky model represents a unique scenario wherein, even when the slow roll conditions are violated, the background, the perturbations as well as the corresponding two and three point correlation functions can be evaluated completely analytically. As a consequence, the Starobinsky model can also be used to calibrate numerical codes aimed at computing the non-Gaussianities.