We do not advise on AI. We build the systems it runs on.

Throughput on autoregressive decoding is bounded by memory bandwidth, not compute. The fix is to stop generating one token per forward pass. We work in speculative decoding — draft candidates cheaply, verify them in a single pass of the target model, keep what survives.

Our implementation is a Medusa-derived multi-head arrangement: additional decoding heads predict several positions ahead, candidate continuations are laid out as a tree, and verification happens in one batched forward pass under a modified attention mask. The published method is a starting point. What actually determines whether you get a usable speed-up is head count and depth against the target model, how the candidate tree is pruned, the acceptance criterion, and how the heads are trained on the distribution the deployment actually sees. That tuning is ours and it stays ours.

The number that matters is accepted tokens per forward pass. Raw tokens per second is a marketing figure.

Full fine-tunes are expensive, destructive to base capability, and immediately stale. We fit parameter-efficient adaptation layers instead — per user, per team, per domain — composed onto the base model at load rather than baked into it.

Layers are versioned, inspectable and reversible. A trading desk, a claims unit and a platform team each get a system tuned to their vocabulary, their risk posture and their tools without four separate deployments and without four separate training budgets. When the base model is replaced, the adaptation is refitted, not rewritten. Adaptation stays an engineering artifact under change control rather than an opaque drift nobody can audit.

If you cannot diff it, roll it back, and say what changed, it is not adaptation. It is damage.

Mixture-of-experts selection at the request boundary, not inside the model. A classifier decides the smallest sufficient model for each request — local or frontier — against cost, latency, sensitivity and jurisdiction.

Most enterprise AI spend is a frontier model answering questions a 7B parameter model on your own hardware would have answered correctly. Routing recovers that without conceding capability on the hard work, and it is where sovereignty actually gets enforced: a request touching regulated material is routed in-perimeter as a policy decision, not as a code change.

Roughly half of observed frontier spend in the estates we have instrumented never needed a frontier call.

Cosine similarity over text chunks retrieves passages that resemble the question. That is not the same as retrieving the passage that answers it, and in a codebase it is frequently worse than grep.

We run a structural index alongside the embedding store — a graph over code, documents and decisions queried by relationship rather than by keyword or distance. Traversal answers dependency questions correctly because it follows real edges. Embeddings handle the fuzzy end. The combination is what makes retrieval survive contact with a large estate, where the naive approach degrades exactly as the corpus gets valuable.

Relationship queries beat similarity queries on anything with structure. Code has structure.

A transformer is the wrong tool for a large share of enterprise problems, and reaching for one anyway is the most expensive mistake in the current cycle. Much of our applied work uses nothing of the kind.

Time-series forecasting and anomaly detection on operational and financial telemetry. Gradient-boosted trees for tabular classification, where they still beat deep models on most real datasets. Computer vision and OCR pipelines for document and forensic work. Signal processing across sensor networks, including indoor localization by three-axis Delaunay triangulation on commodity hardware. Graph algorithms for dependency and fraud topologies. Federated approaches where the training data legally cannot be centralized.

Ask what the model is for before asking which model. It eliminates most candidates immediately.

Sovereign deployment is a hardware problem before it is a policy problem. Getting useful inference out of on-premise equipment is quantization, KV-cache management, batching strategy and honest measurement.

We quantize, evaluate the degradation rather than assuming it, size the KV cache against real concurrency, and tune batching for the arrival pattern the deployment actually has. Then we measure it against the frontier baseline on the client’s own tasks. Sometimes the answer is that local is fine. Sometimes it is that a subset must route out and the policy has to accommodate it. Both are legitimate; guessing is not.

Quantization degradation is task-specific. Benchmark scores will not tell you whether it broke your workload.

You cannot improve what you do not measure, and public benchmarks measure a distribution that is not yours. Every deployment gets a harness built against the client’s own tasks and their own definition of correct.

Regression suites run on model change, adapter change, routing change and prompt change, because all four are silent failure modes. Behavioral drift is caught by the harness rather than by a user complaint six weeks later. This is unglamorous and it is the single highest-value thing most enterprise AI programs are missing.

A model swap without a regression suite is a production change with no rollback signal.