Harnessing LLMs with Deterministic Program Analysis for Legacy Code
There is an optimistic and a pessimistic reading of what GenAI does to the problem of large legacy codebases.
Total Internal Reflection
Technology and Art
Latest posts
Engineering Log: An AI Harness for RedDragon
RedDragon is a multi-language interpreter/compiler that lowers 16 languages (Python, Java, Go, TypeScript, Rust, C, C++, C#, JavaScript, Kotlin, Lua, PHP, Ruby, Swift, Scala, COBOL) down to a shared IR. With 13,000+ tests across 16 frontends, keeping quality high and coverage visible requires a harness that operates at several layers simultaneously. This post documents that harness in full: how…
Engineering Log: Testing agentic coding workflows via Large Refactorings
Two agentic development frameworks applied to the same multi-layer type migration across a 13,000-test compiler pipeline. The first provided process discipline but stalled partway through. The second completed the work because it planned more thoroughly and stopped to ask.
Engineering Log: Anatomy of a Refactoring Using AI
Tracing the full arc of a multi-phase refactoring — from “Java string concatenation crashes the VM” to “every value in the system carries its type” — done across a dozen sessions with Claude Code over two days.
Engineering Log: Building Non-Trivial Systems with an AI Coding Assistant
Notes from building a multi-language code analysis engine across 400+ conversation sessions with Claude Code.
RedDragon: A Deterministic Compiler Pipeline with LLM-Assisted Repair for Analysing Incomplete Code and Unknown Languages
A universal IR with per-opcode typed instructions, 15 deterministic frontends, LLM-assisted repair/lowering/execution, a deterministic VM with class hierarchy support, overload resolution, and cross-language slicing, a structured type system with generics/unions/variance/traits and interface-aware inference, and iterative dataflow analysis.
Building a simple Virtual Machine in Prolog
In this post, I’ll talk about how I wrote a small Virtual Machine in Prolog which can both interpret concrete assembly language-like programs, and run basic symbolic executions, which are useful in data flow analyses of programs. The full code is available in this repository.
Automated Reasoning: A brief overview of Prolog
In this post, I present an abbreviated overview of Prolog and the paradigm of Logic Programming. I’ll discuss why I think it makes for such a powerful domain modelling language (with examples), and a gateway into the techniques of automated symbolic reasoning.
Building an HLASM grammar for reverse engineering, from scratch
This post talks about a technique to build an ANTLR grammar for HLASM (mainframe assembler) from scratch, without writing the grammar of the entire instruction set by hand. The technique creates a parser which reads a table of instruction formats from IBM’s official documentation, and automates the creation of the actual HLASM grammar based on these instruction formats.
Inductor: Automated Hypothesis Verification using LLMs and Hierarchical Bayes-like models
We look at how a Hierarchical Bayes-like model can be used to recursively decompose a hypothesis into sub-hypotheses to form an inference tree. The beliefs of these sub-hypotheses are updated based on the strength of the evidence gathered using MCP tools.