Nirdosha — a systems language designed so your coding agents can ship backend services without human review of every line

I’m reaching out because DeepSeek is among the few labs actually pushing the frontier on agentic coding . You’ve already shown strong results on long-horizon coding, tool use, and agent loops. The remaining bottleneck is no longer “can the model write code?” — it’s “can we trust the code an unsupervised agent writes and maintains?” Nirdosha (निर्दोष — “without fault”) is a research systems language built for exactly that problem. The core idea A language an AI agent can write and that the runtime can trust without a human reviewing every line. Key properties (already implemented in a real Rust-based compiler): LL(1) grammar exported to GBNF → constrained decoding forces every token the model emits to stay syntactically valid. No data races, no deadlocks by construction (no mutex primitive exists in the language). Affine ownership + refinement types → no GC, no integer/buffer overflow in accepted programs. Structured diagnostics (JSON) so an agent gets a machine-readable proof obligation instead of a paragraph of English to guess at. Real OS-process sandbox as a language primitive, not a bolted-on Docker wrapper. Capability / role gating at the call site (requires(role: "admin") etc.). UI engine that derives live CRUD dashboards + role-aware screens from ordinary structs. It is deliberately not trying to be a better Rust or a general-purpose language. It targets one unsolved slice: backend services and compliance CRUD written and maintained by coding agents . Status: active research prototype. Compiler is real and runnable; many safety properties are proven today; remaining gaps are honestly documented in the wiki. Repo + wiki: github.com/arunsoman/nirdosha (Especially: Design Philosophy, LLM Integration, Honest Scope & Roadmap) Why this is relevant to you Your models already generate and repair code at high volume. The missing piece is a language whose grammar and type system make whole classes of failure unrepresentable , so an agent can be constrained to stay inside the safe subset and still produce useful systems software. If Nirdosha reaches its potential, it becomes infrastructure that multiplies the reliability of any strong coding model — including yours. What I’m asking I’m looking for partners who care about the agent-written-systems problem and can help accelerate the remaining engineering (codegen completeness, real-LLM evaluation harness, performance, ecosystem). Possible forms of support (happy to discuss which fits best): Compute / research collaboration on agent evaluation with your models Engineering support or joint work on the compiler / constrained decoding path Visibility or integration experiments (e.g., Nirdosha as a first-class target for your coding agents / CLI tools) Advisory or light sponsorship to keep the work moving I’m happy to jump on a short call, walk through the current capabilities and the honest gaps, or send a more detailed technical brief. Thanks for the work you’re already doing on the agentic side — it makes projects like this finally practical. Best regards,