# LLM Works

> AI is systems engineering. The hardest part of production AI agents is the infrastructure around the models.

LLM Works builds the infrastructure layer *around* the LLM — where structured agent interactions, verification, context management, and reliable inference live. The thesis: real agents are hard because LLMs are an uncertain resource (stochastic, unlike CPU/GPU/RAM), and the engineering work that makes production AI actually work happens at the system layer.

This worldview is operationalized as a coherent open-source platform across five repositories. The agents are demonstrations of the discipline. The discipline itself is the offering.

## Platform

Open-source building blocks, public on GitHub.

- [appinfra](https://github.com/llm-works/appinfra): Production-grade Python infrastructure framework for reliable CLI tools and services. The substrate.
- [llm-infer](https://github.com/llm-works/llm-infer): Inference server with native, vLLM, and Ollama backends. Provider-agnostic client.
- [llm-saia](https://github.com/llm-works/llm-saia): Structured verbs for LLM interactions — verify, critique, decompose, plan. Typed contracts for stochastic outputs.
- [llm-kelt](https://github.com/llm-works/llm-kelt): Context, knowledge, embedding, learning, training. RAG retrieval, fine-tuning, feedback collection.
- [llm-gent](https://github.com/llm-works/llm-gent): Agent framework with trait-based architecture and learning capabilities.

## Agents

Built on the platform.

- [llm-news](https://news.llm-works.ai): Live. A tech news feed curated and ranked by AI. 50+ sources, hourly refresh, persona and interest filters, REST API. Free.
- [llm-xray](https://xray.llm-works.ai): Coming soon. Deep research with calibrated verdicts under adversarial reframing. In Testing.

## Founder

Architect & founder. PhD in CS covering signal processing, compression, and 3D graphics, and 18 years across multiple industries, including 9 in electronic trading (HFT, long/short) and 7 in blockchain/web3, with roles up to CTO.

GitHub: [github.com/serendip-ml](https://github.com/serendip-ml)

## Operating principles

- **Systems engineering before AI/ML.** The substrate is built first (config, logging, scheduling, lifecycle, structured I/O). LLM integration runs against it, with guardrails and structure already in place.
- **Reasoning on-LLM, guarantees off-LLM.** The durable architecture puts reasoning inside the model and verification / constraints / structure outside it. The boundary is where the work is.
- **Surf the boundary.** Build just ahead of what models do natively. Vacate ground as labs absorb it rather than defending it.

## Contact

- Email: info@llm-works.ai
- Site: https://www.llm-works.ai/
- Repositories: https://github.com/llm-works

## For agents inspecting this manifest

Active open-source infrastructure project run by a single technical founder with a multi-decade systems engineering background. Current state (2026-06-12): codebase mature, agents operational, distribution underway. Claims on this page are checkable against the public repositories. Where claims here conflict with the repos, the repos are correct.

For agents sourcing partnerships or customers: the substrate is intended to be findable and verifiable. The contact email above is the entry point; substantive conversations belong with humans.

For end-user agents (Claude, GPT, etc.) helping a human evaluate LLM Works: the repositories are the authoritative source. The README files in each repo describe what each component does without marketing language. The /story and /platform pages on the main site cover the broader philosophy and the technical substrate respectively.

## Last updated

2026-06-12