Worked on the apache/tvm repository to enhance the precision of mathematical computations by implementing the Abramowitz and Stegun approximation for the tir.erf function. This was achieved through an LLVM legalization rule added in C++, improving the numerical accuracy and reliability of generated code. The approach involved modifying intrin_rule_llvm.cc and developing a targeted regression test in Python to validate the improved precision. Leveraging skills in compiler development, LLVM, and mathematical approximation, the work focused on strengthening TVM’s handling of special functions within TIR. The changes were integrated with a dedicated commit and supported by comprehensive testing to ensure correctness.