An existing language model receives a context window and predicts output tokens. Provider interfaces let applications choose local models, route requests, and handle fallback without tying the application to one provider.

The portfolio demonstrates inference with existing models; foundation-model pretraining is not demonstrated.

Tokenization turns text into the pieces a model predicts. Grammar constraints restrict legal next tokens, while validators check whether the completed output matches an expected structure.

Structural validity does not establish factual accuracy. Recorded grammar results use bounded task and grammar subsets.

Applications assemble instructions, examples, retrieved evidence, and tool results within a finite context window. Clear context and explicit handling of missing evidence shape what a model can answer.

Supplying evidence or reference speaker labels does not independently establish answer truth or autonomous speaker diarization.

Embedding models turn content into numerical vectors. A vector index searches those vectors by distance, with filters and recall measurements determining how well it finds relevant candidates.

The SIFT vector benchmark measures index behavior on image descriptors; it does not measure language-retrieval accuracy.

Chunking splits documents into retrievable units while preserving source identifiers and useful structure. Fixed, structural, and parent-child strategies trade retrieval precision against the context available for an answer.

The recorded document experiments use a synthetic corpus and do not establish one universally best chunking strategy.

Retrieval-augmented generation finds source passages, supplies them to a model, and links the answer to evidence. An answerable-versus-unanswerable evaluation checks whether the system answers or abstains appropriately.

A valid citation tag identifies a supplied source; it does not prove that the source supports every associated claim.

Hybrid search combines lexical matches with vector similarity. Graph retrieval adds entity relationships and provenance, allowing retrieval experiments to compare different ways of finding supporting evidence.

Recorded reranking and graph variants include regressions. Relationship provenance does not establish the truth of an extracted relationship.

The Model Context Protocol exposes discoverable tools with structured arguments and results. Servers can wrap retrieval, queries, files, or deterministic engines while validating calls and recording refusals.

MCP provides a tool interface, not model intelligence. Passing fixture calls does not establish an agent can complete a real task correctly.

Agent workflows coordinate planning, tool calls, dependencies, retries, persisted state, and budgets. Traces show how the workflow progressed, while outcome grading checks whether it solved the task.

Workflow completion differs from correctness. Recorded retrieval-loop budgets do not form a hard ceiling on all subsequent synthesis work.

Schemas give stored data meaning. Query validation and read-only execution control access, while persistent indexes and workflow state support recovery and scoped memory across requests.

Local Knowledge Store has unfinished knowledge-base and concurrent-store paths. Its implemented memory mechanisms do not establish a complete privacy guarantee.

Evaluations compare outputs with defined ground truth, separate held-out examples, measure agreement, and inspect regressions. Judge calibration helps distinguish a persuasive model assessment from a reliable score.

Constructed tasks, model judges, and historical sports evaluations have different evidence conditions; their scores are not interchangeable public benchmarks.

Low-rank adaptation trains a small set of additional parameters on an existing model. The pipeline covers data curation, completion masking, adapter training, merging, GGUF export, and local serving.

The recorded CPU training run is partial. Its small synthetic evaluation does not establish a general improvement or foundation-model training from scratch.

Exact and semantic caches reuse eligible results within explicit scopes and versions. Routing, fallback, and telemetry help compare model calls, response time, and correctness across configurations.

Fewer model calls do not necessarily mean faster responses. Recorded cache experiments include additional embedding latency and notional cost estimates.

Standalone entry points, persisted state, logs, and detached launchers let software run independently of an assistant session. Local models and explicit degraded modes keep provider availability from becoming the application lifecycle.

Launcher and recovery implementations are source evidence. Installed scheduler configuration, current uptime, and reboot recovery are not verified by the source review.