You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
WorkstationLLM is an open-source defensive primitives toolkit for operators running local-LLM stacks on consumer hardware. Three independent daemons (memory guard, request queue proxy, disk watchdog) prevent the kernel-panic, OOM, tool-spiral, and disk-fill classes of failure that take down hosts running production AI workloads. The toolkit has supervised my own 24-application local stack on a single Mac Mini M4 Pro crash-free for over 6 months. Apache 2.0 from day one. This is a distinct project from my existing Manifund posting NoEnterprise (the 5-agent C-level autonomous organization). NoEnterprise is the agent organization that runs on top; WorkstationLLM is the safety substrate underneath. The two complement each other and target different deliverables.
Goal: ship WorkstationLLM as a public open-source release that any local-LLM operator can adopt to prevent the failure modes that kill hosts running consumer-hardware AI workloads. Five concrete deliverables in 6 weeks from minimum funding: (1) One-binary installer for Mac, Linux (Ubuntu 24.04 with CUDA), and Windows (WSL). (2) Conformance test suite with 12 to 15 named defensive-readiness test cases, OS-agnostic Python harness, pass/fail criteria, and scoring rubric. Apache 2.0. (3) Operator handbook (CC-BY 4.0) covering threat model, threshold tuning, integration patterns for vllm-mlx, mlx-lm, Ollama, and Llamafile. (4) Reference report from my own production stack as the canonical baseline (24 apps, 6+ months, no panics). (5) Community rollout: Apollo Research Slack, METR mailing list, Apart Research Discord, MATS alumni post, plus three long-form case-study blog posts. How will you achieve them: I have shipped the Mac-native version already. The toolkit runs my own stack today. The funding pays for focused engineering time on the Linux + CUDA port (the hard part, since memory-pressure APIs and disk-watch primitives differ across OSes), the security audit, and the community rollout work.
Total budget $30,000 USD, fundable in pieces by multiple regrantors. Specifically: $12,000 for code refactoring (Mac to Linux cross-platform, the harder of the two OS bridges). $8,000 for the conformance test suite implementation. $6,000 for documentation and operator handbook. $4,000 for community rollout and three case-study blog posts. I will accept partial funding gratefully. The $5,000 minimum unlocks the cross-platform port (deliverable 1). Full $30,000 ships all five deliverables in 6 weeks.
Solo. Me. Ryan Hammer, public name NoBanks Nearby. 40 years old, Apple Valley California. Late-diagnosed Asperger's plus Bipolar II at 40. 17 to 18 years as a working cinematographer since 2008 (mixed roster including some Fortune 500 brands alongside many other companies and independent projects). Self-taught engineer since July 2024 (22 months). 24 production AI applications shipped on a single Mac Mini M4 Pro with no cloud bills. Track record on similar safety/infrastructure work: WorkstationLLM has supervised my own 24-application local stack crash-free for over 6 months. The toolkit was built because the panics WERE happening; every primitive is a response to a specific failure mode I lived through. I am the upstream contributor to the Hermes agent framework (patches in v0.13.0 The Tenacity Release related to worker lifecycle, zombie detection, and retry-cap fixes). Other shipped work: ACiD (a public Ethereum L2 at acid.livingagentic.me, Chain ID 1714 verified live), TRAIDE (a 22-chain non-custodial DEX with 231 contracts deployed at deterministic CREATE2 addresses), MAITE (an AI personal coach in 8 languages), TINP (a generative-art pipeline that has produced original art every single day for 218+ consecutive days at thisisnotphotography.com), and three SaaS products (REDLINE for legal contract review, DATAROOM for M&A due diligence, COVENANT for portfolio monitoring). Existing prior project on Manifund: NoEnterprise (the 5-agent C-level autonomous organization). NoEnterprise is the agent org; WorkstationLLM is the substrate. Distinct projects, complementary outputs. GitHub: https://github.com/NoBanks LinkedIn: https://www.linkedin.com/in/nobanksnearby/ X: @livingagentic
Most likely failure cause: the Linux/CUDA port surfaces a defensive-primitive that does not have a clean cross-platform analogue. For example, Mac's memory_pressure tool has no exact Linux equivalent; the conformance suite would have to define an equivalent threshold metric via /proc/pressure on Linux kernel 5.2+ or fall back to vm.stat polling. If I cannot land that bridge in 6 weeks, the Linux deliverable slips to 12 weeks. Partial funding outcome in that scenario: Mac binary plus conformance test suite ship on time; Linux ships in v0.2 instead of v0.1. Second failure cause: community rollout under-reaches because the toolkit is too niche. Mitigation: the AI safety research communities (Apollo, METR, Apart, MATS) all run local inference today, so the audience exists. If the rollout under-performs, the Apache 2.0 release stays online indefinitely and benefits whoever finds it. Third failure cause: my health. Bipolar II plus Asperger's means I have bad weeks. Mitigation: 22 months of consistent daily shipping is the track record. The toolkit got built during the bad weeks too; the system runs on documentation and primitives, not heroics.
Zero. Pre-revenue, pre-incorporation by design. The 24-application stack was built on a $2,300 Mac Mini bought with personal savings, no cloud bills, no outside funding. Every artifact is bootstrapped. I have submitted approximately 52 grant applications across the last 90 days targeting open-source AI safety, sovereign-AI infrastructure, multi-agent systems, and adjacent categories. A small number have been rejected (Emergent Ventures, BlueDot withdrawn for integrity reasons), most are still in review. Total submitted ask across all open applications is in the $1.5M to $15M range; expected hit rate is low single digits, so realistic expected value is in the low-five-figure to low-six-figure range across the next quarter. Manifund regrantors who have already funded NoEnterprise on this platform are the most-likely first funders of WorkstationLLM because the substrate is what makes the agent org work.