Skelf projects vs. the alternatives.

Prompt optimisation

Angle: Declarative prompt specification vs. programmatic prompt composition

promptel is a specification language and tooling; DSPy is a Python framework for compiling prompts. promptel optimises for portability, version control, and human review; DSPy optimises for runtime bootstrapping. They are complementary, not competitors — promptel can express a DSPy signature, and DSPy can compile a promptel prompt into a teleprompter.

Agent memory

Angle: Structured persistent memory vs. mostly-vector memory

memorg is opinionated about structured schemas (semantic facts, episodic traces, procedural skills) rather than treating memory as a single vector store. If you need a quick RAG-memory layer, Mem0 / Zep are faster to start. If you need queryable, auditable, schema-validated memory for long-running agents, memorg fits.

Agent protocols

Angle: Compliance + audit layer on top of MCP and A2A

mpl is not a replacement for MCP or A2A; it is a layer that sits between your agents and the underlying transport. Where MCP and A2A define how agents talk, mpl defines what correct looks like — typed contracts, quality metrics, cryptographic audit trails, and policy enforcement. mpl maps directly to SOX, GDPR, HIPAA, and the EU AI Act.

Local LLM serving

Angle: Research-instrumentable local LLM server vs. production-ready alternatives

Ollama is the default for getting started; vLLM is the production-grade GPU server; LM Studio is the desktop GUI. mullama is the research-focused alternative that exposes llama.cpp internals for instrumentation — useful when you need to study scheduling, KV cache, or model lifecycle as research artefacts rather than treating them as black boxes.

LLM routing

Angle: MIPROv2-based auto-tuning vs. hand-coded routing

LiteLLM, Portkey, and OpenRouter are gateways; route-switch is a router that learns. It uses MIPROv2 to automatically tune per-model prompts against your traffic, optimising the cost-quality frontier empirically rather than relying on rules.

Sandboxing

Angle: Minimal-overhead sandboxing in pure Zig vs. heavier alternatives

gVisor is a mature, heavy userspace kernel; Firecracker is a microVM; WASM is portable but constrained. zviz is a gVisor-inspired sandbox in pure Zig that targets near-zero runtime overhead — the right choice when you need to run many short-lived untrusted code executions per second without paying the per-execution cost of a full VM.

Vector search

Angle: Embedded SQLite-backed ANN vs. dedicated vector databases

For sub-100K-vector corpora and prototype workloads, memista runs inside the same process as your application, with no separate service to deploy. For million-vector corpora with high QPS, Pinecone, Qdrant, or Milvus are better choices. memista is the right answer to "do I really need a vector database?"

Programmable databases

Angle: Embedded Lua-scriptable database vs. SQL-only embedded DBs

SQLite and DuckDB are the gold standard for embedded SQL, but they push logic out to the application. liath treats the scripting layer as a first-class citizen: queries and transformations are written in the same language as the application code. liath-rs adds a Rust reimplementation with RocksDB for performance comparison.

NUMA-aware scheduling

Angle: NUMA-first Rust crate for AI workloads vs. general-purpose tools

hwloc and libnuma are the general-purpose NUMA libraries; numaperf is opinionated about AI workloads and the specific topology patterns they exhibit (KV cache locality, attention memory patterns, batch-aware pinning). It measures and reports the latency impact so you can defend the topology-aware decision with numbers.

NL to constraint satisfaction

Angle: LLM-to-formal-solver pipeline vs. pure LLM output

Pure LLM output cannot guarantee optimality or even feasibility. savanty uses an LLM to translate English problem descriptions into a formal solver input (Z3, OR-Tools, MiniZinc) and delegates the actual solving to the formal solver. The output inherits the solver's mathematical guarantees.

Ranking with sparse feedback

Angle: Multi-armed bandit ranking vs. classical ranking models

Bradley-Terry and TrueSkill need many comparisons to converge. compere uses multi-armed bandit algorithms to find the correct ordering with O(n log n) comparisons instead of O(n²). For tournament design, search evaluation, and recommendation feedback, that difference is the difference between feasible and infeasible.

Trading signal compilers

Angle: Visual-to-Rust compiler for quant signals vs. algorithmic DSLs

QuantConnect and Lean are full algorithmic trading platforms with their own DSLs. sigc is narrower and faster: it compiles a visual signal specification directly into verified Rust executables. The thesis is that most quant alpha is in the signal definition, not the surrounding platform — sigc is for the signal-definition-first workflow.

On-device Flutter LLM

Angle: On-device LLM runtime for Flutter vs. RN bridges

llama.rn is the React Native equivalent; flutter_llama_cpp is the community Flutter binding. llamafu is the research instrumented version that publishes its measurements — token/s per device, memory ceilings, quantisation quality trade-offs — and is built on a measurement-first methodology.

On-device mobile AI agents

Angle: On-device autonomous mobile agents vs. cloud-based agents

ChatGPT Operator and Anthropic Computer Use are powerful but require sending the user's screen contents to a remote server. ukkin is fully on-device: the agent runs on the phone, never sends data to the cloud, and operates under a tiered permission model that limits what an agent can do without confirmation.

Deliberative search

Angle: Reasoning before retrieval vs. retrieval-then-ranking

Traditional search retrieves documents that match the query and ranks them. slorg uses a language model to reason about query intent first, then issues targeted retrievals. For ambiguous or complex queries, this dramatically improves precision. The term is sometimes used interchangeably with "agentic search".

Browser-extension LLM frameworks

Angle: Open framework for LLM browser extensions vs. vendor SDKs

Most LLM browser extensions lock you to a single provider. anouk is a portable framework that abstracts the manifest v3 quirks and provides a unified API for the LLM call — you can swap OpenAI, Anthropic, or local llama.cpp without rewriting the extension.

Ephemeral credentials

Angle: Ephemeral credential proxy for LLM APIs vs. general-purpose key management

General-purpose API gateways and secret managers do not understand LLM-specific scopes (model, max tokens, allowed endpoints). perishable is purpose-built for LLM access: it issues tokens that are scoped to a specific model, with a max token budget, and a TTL measured in minutes.

Warehouse robotics simulation

Angle: Deterministic RL-ready RMFS simulator vs. legacy simulators

RAWSim-O and ARENA-Sim are the legacy RMFS simulators. waremax is built around three properties they do not jointly provide: exact determinism (byte-identical replay), a first-class RL interface (Gymnasium + PyO3), and instrumented delay attribution usable as a reward signal.

Independent AI research labs

Angle: Independent research lab publishing production-grade OSS as primary output

Big Tech AI labs publish papers; AI startups publish products; Skelf Research publishes runnable, testable, peer-reviewable code. The methodology — "hypotheses as software" — is the differentiator. We are not competing for benchmarks; we are building a public corpus of falsifiable AI research.