Pinull

Problem: AI coding assistants are fast, but they invent functions that do not exist, misuse libraries and produce code that looks right and is not. Developers must re-check everything, and in banking, government or healthcare one unnoticed error costs money, data or trust. Solution: Pinull, built by Krydia, puts a deterministic verification layer between the AI model and the user. Before any code is written, the user reads and approves a machine-checkable specification: signature, examples, edge cases, error behaviour. The model then writes the code, and four gates check it: every import, name, attribute and call must exist in the real interpreter that will run it, types must hold, data flow must be safe (path traversal, SQL injection, leaked secrets), and quality rules must pass. Tests are generated from the approved specification, not by the model that wrote the code, run in isolation, required to reach 90% of lines and branches, and measured by breaking the file on purpose. Only then is the file handed over, with a passport of what was proven. Otherwise Pinull refuses and names what is still undecided. Never both. Audience: software teams adopting AI coding, especially in regulated sectors, and developers who already use Claude, ChatGPT or Gemini: Pinull runs inside them as an MCP server, on the user's own model. Status: working MVP for Python, 858 automated tests, every public guarantee bound to a test that enforces it. Early access is open at pinull.com. Why it deserves the nomination: it attacks the core weakness of generative AI, confident fabrication, with measurement instead of promises.

Nomination

AI Startup of the Year

Company

Krydia

Project

Pinull

Sector

Cybersecurity / кибербезопасность

Stage

Идея / прототип

AI component

Pinull is a neurosymbolic system: large language models generate, deterministic and symbolic methods decide. 1. LLMs (any OpenAI-compatible provider, with automatic failover between Gemini, Groq and Mistral) write the specification and the code, and repair it from exact problem reports. 2. A verification engine resolves every reference with Python's own scoping rules and checks it against the live target interpreter: imports, attributes, call signatures. 3. Strict static type analysis, plus our own data-flow analyzer that follows values through assignments, strings, containers and helper functions to catch path traversal, SQL injection and leaked secrets. 4. Property-based and differential testing: 200 generated inputs per property, and comparison with a standard-library reference where one exists. 5. Coverage and mutation testing prove the tests themselves could fail. 6. An agentic repair loop: each rejected attempt goes back to the model with the precise reasons, within a fixed budget. Researched and documented as next steps: symbolic execution (CrossHair, Z3) and formal proof (Dafny). Human role: the user approves the specification before any code exists, the one decision a machine cannot make for them, and every claim in the result states its evidence and the interpreter it was checked against.

Impact / results

Every number is measured and reproducible from our repository: - Hallucination audit: 432 faults injected into real code, and not one invented reference came back verified. 82.7% were refuted with the exact reason, the rest marked unverifiable, never passed. - False alarms on working code: 0.16 per 1,000 references across the Python standard library. 81.9% of all references are decided automatically. - 53-task benchmark with independent behaviour checks (latest run, 51 tasks measured): 32 of the 34 files Pinull accepted were correct. In both exceptions the model misread the task in the specification the user approves before code is written. The rest were refused with reasons, not handed over unproven. - Security: on 23 purpose-written vulnerable files, standard linters caught 0 path traversals, 0 secret leaks and 2 of 5 SQL injections, the gaps our data-flow gate closes. Run over 4,449 real-world files with 0 crashes. - Real bugs found in widely used software while auditing it: a path traversal in CPython's pydoc web server, a missing import in CPython's IDLE, a module used without import in httpcore. - 858 automated tests. 294 links bind every public guarantee to a test, and CI fails if one breaks.

Innovation

Most AI coding tools optimise for how much code they produce and leave checking to the developer. Pinull inverts this: the model is an untrusted proposer, and nothing reaches the user unless deterministic checks prove it. 1. Verified against the actual interpreter and packages that will run the code, not against the model's memory of an API. 2. Specification first: the user approves a machine-checkable specification before code exists, and the tests are generated from it, so the model never grades its own work. 3. Tests that are themselves tested: 90% line and branch coverage and mutation strength. A suite that could not fail does not count. 4. Honest refusal as a feature: "I could not prove this" is an answer, never a silently wrong file. 5. A public, test-enforced ledger of guarantees: every promise names the test that keeps it. Pinull does not compete with AI assistants. It makes them trustworthy, working inside the tools developers already use.

Connection to Azerbaijan

Pinull is designed and built in Azerbaijan by Krydia. It serves a goal of the Artificial Intelligence Strategy of the Republic of Azerbaijan for 2025-2028: safe, trustworthy adoption of AI. Our first target users are Azerbaijani software teams, banks, fintechs and public-sector IT units that want the speed of AI coding without the risk of fabricated code in critical systems. Next steps: pilots with local companies, publishing our benchmark and method as an open reference for the local AI community, workshops for universities and bootcamps on verifying AI-written code, and specifications and reports in Azerbaijani. Pinull is built for the global market of AI-assisted software development, and Azerbaijan is where it starts. Every line proven by Pinull is a line Azerbaijani engineers can trust.

Team

1–5 человек

Business model

Другое

Launch date

2026-10

← All nominees