Investor Presentation · June 2026
semiring.ai
Deriving efficient and reliable code for AI and scientific computing.
AI driven software engineering infrastructure translating declarative specfications into optimised code step by step, recording every transformation, assumption and decision.
The Problem
There is no straightforward path from a mathematical model to production-ready code.
Engineering and research effort is required to work around the realities of computers:
Traditional compilers only address low-level details while algorithm design still happens at human pace in research papers.
  • Numerical stability
  • Approximations
  • Algorithm choice
  • Data movement
  • Scheduling
  • and more…
AI Coding
AI coding agents are powerful but results for kernel optimisation and scientific programming are mixed.
Reliability
Agents can propose program transformations that go beyond any traditional compiler optimisations, but the proposals are unreliable. Reliability issues compound even more when giving the model a memory to record lessons learned.
Capability
AI spots many micro-optimisations but often misses the big picture. The most impactful improvements require taking a step back and making algorithmic improvements.
Capability and reliability reinforce each other: better correctness infrastructure lets AI be bolder, and better verified transformation tools make AI more dependable.
Our Insight
We don't hand the AI an existing implementation and ask it to improve it. By that point, the original mathematical intent is buried under accumulated choices - there's nothing left to reason about. Instead, our process is top down:
I
Start with the math
f(x) = ‖Ax − b‖² DECLARATIVE SPEC
We begin with a specification of the problem - a mathematical model that captures exactly what needs to be computed, independent of how.
II
Lower step by step
Spec IR₁ IR₂ Code
We derive code through a sequence of explicit choices - algorithm selection, approximations, scheduling. Each step is recorded in the trace alongside the assumption it introduces.
III
Test
Spec IR₁ IR₂ Code
We test that the output faithfully implements the original spec and that every assumption holds. When something fails, the trace points to the exact decision that introduced it.
IV
Iterate
Spec IR₁ IR₂ IR₂’ Code
The data gathered during testing is used to identify improvements. Failing decisions are revisited at the exact point they were made. New candidates - from the AI, the engineer, or automated search - are proposed in their place.
The Product
Three layers that turn mathematical intent into deployable code
A complete solution has to span mathematical semantics, compiler engineering, and AI-guided search.
i
Unified Representation
A program representation that can capture the mathematical spec, the final implementation, and the partially lowered stages in between in one coherent object. This cannot be plain source code.
ii
Semantic Compiler
The compiler provides powerful transformations over that representation. Trusted transforms preserve correctness by construction, so the system can do more without re-verifying every step from scratch.
iii
Derivation Harness
Optimization is an experimental loop. The harness runs candidates, collects outputs and profiling data, and compares alternatives statistically to drive the next iteration.
Why Now
AI Demands a New Compiler Stack
AI opened a new design space for compilers
Before useful models, most compiler intelligence had to be encoded manually as heuristics. AI makes it possible to build compiler infrastructure that can search, reason, and adapt in ways that were previously impractical.
AI engineering works better with guidance
AI systems can produce useful engineering work, but results improve sharply when they operate inside explicit structure. Feedback, evaluation, and constrained workflows make the system more effective and more dependable.
Computation now drives science and business
More outcomes now depend on computational techniques, and those outcomes scale with more compute. That creates a large opportunity for infrastructure that helps engineers build software that is more performant, more reliable, or more capable.
Market & Go-to-Market
Land one vertical. Expand as a platform.
01
Land
One vertical where fast-and-defensibly-correct commands a premium: HPC simulation, quantitative finance, or ML systems work. Sharpest pain, clearest buyer, strongest willingness to pay. Narrow and deep earns the right to generalize.
02
Expand
As more domains contribute dialects and notions of evidence, platform value compounds. Each new vertical adds reusable fidelity theories that reduce the cost of the next one. The substrate appreciates with scale.
03
Buyer
Teams whose results must be both performant and defensible: research labs, financial systems, regulated-domain ML, and engineering teams already doing expensive expert-driven optimization by hand.