ISP RAS · SharpChecker
Symbolic computation and structured C# code/declaration metadata for static analysis and AI-assisted defect recognition.
Professional profile / 03 — 03
I build compiler/runtime components in C#/.NET and C++/LLVM, from optimization-legality checks to generated-code validation. At ISP RAS I work on C# symbolic analysis in SharpChecker; at MCST I worked on LLVM/LICM. I keep correctness evidence and performance measurements separate for experimental optimizations.
Commercial work, research and independent projects are presented separately, without inflating tenure.
Symbolic computation and structured C# code/declaration metadata for static analysis and AI-assisted defect recognition.
C++23/LLVM 22: experimental LICM with conservative legality checks; Python differential checks of before/after behavior.
Application and backend work with C#/.NET, Unity, ASP.NET Core, SQL/PostgreSQL, file I/O and servers. Duration is candidate-reported.
Real constraints, engineering choices and verifiable outcomes.
.NET · modular language/toolchain infrastructure · runtime architecture
Module contracts and dependencies are checked in LanguagePlan before execution. Separate interpreter/CIL backends allow result comparisons for the same language, with dedicated parity scenarios.
Frontend, semantic components, IR/optimization passes, and execution backends should evolve independently while composing into one deterministic toolchain without hidden registration-order dependencies.
Typed contracts/capabilities, dependency/conflict resolution, deterministic contribution ordering, and typed artifact routes are resolved into one LanguagePlan before runtime materialization.
Implemented an end-to-end pipeline with frontend, IR, optimization passes, interpreter, and a CIL backend; current exact test manifest: 1,324/1,324.
C++23 · LLVM 22 · interprocedural analysis · optimization
Interprocedural effect analysis is separated from PromotionLegality and the IR transformation. Volatile/atomic accesses, unknown calls and unsupported CFG conservatively reject optimization; negative regression cases cover unsafe paths.
The pass must recognize affine evolution of global IVs across calls and transform IR only when correctness conditions are established.
APInt affine b + k·x evolution, transitive function/module effects, loop/dominator/must-execute reasoning, and a separate conservative legality phase that rejects unsupported CFG/calls.
A conservative LLVM 22 prototype/MVP with an actual IR transform; 29 positive/negative regression cases cover verifier validity, idempotence, and before/after execution.
Rust · compiler backend · register allocation · SysV x86-64
SSA/CFG → liveness → register allocation → x86-64 pipeline. An independent verifier checks register assignments and spills; native execution is compared against a reference IR interpreter.
The goal is to go from IR to executable x86-64 and verify that register allocation preserves semantics.
SSA-like IR with CFG/SSA validation, liveness/interference, register allocation with an independent assignment verifier, phi lowering, and SysV x86-64 code generation.
A reference interpreter and differential execution check generated-code semantics; the allocator assignment contract is verified independently.
Experimental .NET runtime and performance
DynamicMethod/Reflection.Emit and RuntimeMethodHandle; BenchmarkDotNet compares reflection Invoke, delegate and pointer calls on generated methods.
Public .NET abstractions can obscure invocation overhead and the fragile boundary around CLR internals.
Experiments with DynamicMethod, Reflection.Emit, RuntimeMethodHandle and native function pointers; BenchmarkDotNet comparisons of Invoke, delegates and pointers.
A reproducible research prototype, not a production optimization or a general CLR performance claim.
Technologies are targeted to this CV edition; usage context is in the case studies.
MEPhI Junior 2026, 1st-degree diploma; LangDev 2026 speaker.
Aggregate commercial tenure is candidate-reported. Source availability, a test manifest and freshly executed tests are different evidence levels. TuringSquare-LMS is in development: parts are agent-implemented, payments are sandbox only. None of these statements claim independently verified production behavior.
Read source limitationsOpen to conversation / 03
Mikhail Razakov · Moscow / remote