The AI infrastructure layer for BlackForge. Rows are features ordered by dependency inside each workstream; columns are quarters; a chip marks the quarter a feature ships. Hover a row to trace what it needs (upstream) and what it unlocks (downstream). Tap or click a row to pin the trace and open its implementation details. Click a quarter header to show only that quarter's features; click it again to restore.
Tap any feature in the matrix to open its details — how it fits the current stack (bf-* services, DGX fleet) and open-source options to adapt or embed. Printing this page includes every feature's details expanded.
Model class is a first-class registry attribute: it determines the served endpoint, the applicable fine-tuning methods, and the dataset types those methods require. Embedding + reranker + chat compose the RAG stack consumed by Nexus RAG.
| Model class | Endpoint | Fine-tuning methods | Dataset types | Arrives |
|---|---|---|---|---|
| Chat / instruct LLM | /v1/chat/completions | LoRA/QLoRA SFT · DPO · RFT | conversation · instruction-pairs · preference-pairs · graded-tasks | Serving live · FT Q1'27 |
| Base / completion LLM | /v1/completions | Continued pretraining · full-parameter SFT | text-corpus | Gateway Q3'26 · FT Q2'27 |
| Embedding | /v1/embeddings | Contrastive fine-tuning | contrastive-triplets | Gateway Q3'26 · FT Q1'27 |
| Reranker | /v1/rerank | Pairwise / listwise ranking FT | ranking-lists | Gateway Q3'26 · FT Q1'27 |
| Vision LLM | /v1/chat/completions (image parts) | Multimodal SFT | image-text-pairs | Serving Q1'27 · FT later |
Features added by the gap analysis and architecture review, now slotted into both the planning canvas and this matrix (marked NEW in the matrix): /v1/rerank → Q3'26 gateway epic · model intake pipeline, platform observability, request/response logging → Q4'26 · dataset foundation, eval harness v0, vision-model serving → Q1'27 · DPO + full-parameter SFT, quantization-aware training, offline distillation → Q2'27 (multi-LoRA moves here too, off the end of a dependency chain that previously had to complete within one quarter) · RFT/RL, inference guardrails, and fine-tune-from-traffic remain in the vision column.
Deliberately out of scope: audio transcription, image generation, disaggregated prefill/decode, trillion-parameter MoE serving. Differentiator kept: sovereign air-gapped bundle (Q2'27).