Acumen Getting AI to Work Let’s talk
GETTING AI TO WORK / AI ENGINEERING

AI engineering.
The depth behind the outcome.

Useful AI depends on more than an API call. Explore the retrieval, models, infrastructure, speech systems, security controls, and evaluation disciplines that make enterprise AI workable.

/01

Retrieval that earns
the answer.

ACUMEN / ENGINEERING NOTETwo retrieval paths, one evidence setFIG. HUB
Two retrieval paths, one evidence setLexical and vector candidates meet in a filtered, ranked context; the answer layer receives selected evidence rather than the entire repository. Components: Parsed sources; Metadata + versions; Vector index; Exact-term search; Fusion + reranking; Grounded context. ACCESS FILTERS APPLY BEFORE EVIDENCE REACHES THE MODEL 01Parsed sources02Metadata + versions03Vector index04Exact-term search05Fusion + reranking06Grounded context
Lexical and vector candidates meet in a filtered, ranked context; the answer layer receives selected evidence rather than the entire repository.Scroll the drawing sideways to inspect it.

Chunk boundaries, filtering, ranking, and source lifecycle determine whether the model ever sees the evidence it needs.

Read the technical perspective /02

Agents with tools.
Systems with boundaries.

ACUMEN / ENGINEERING NOTEA controlled graph rather than an open-ended loopFIG. HUB
A controlled graph rather than an open-ended loopThe execution path includes a review branch and a recorded return path from every side-effecting tool. Components: Task state; Plan / interpret; Validate arguments; Approval gate; Execute tool; Persist + reconcile. AUTHORIZATION BOUNDARYINTERPRETATION ≠ AUTHORITY 01Task state02Plan / interpret03Validate arguments04Approval gate05Execute tool06Persist + reconcile
The execution path includes a review branch and a recorded return path from every side-effecting tool.Scroll the drawing sideways to inspect it.

Tool design, state transitions, and authorization deserve more attention than the agent’s personality.

Read the technical perspective /03

Adapt the model.
Measure the difference.

ACUMEN / ENGINEERING NOTEBase capability plus a measured adaptationFIG. HUB
Base capability plus a measured adaptationThe training set changes the model; the held-out set determines whether that change is useful. Components: Reviewed dataset; Training split; Base model; Adapter ΔW; Held-out test; Versioned deployment. TRAINING CHANGES WEIGHTSTESTING EARNS RELEASE 01Reviewed dataset02Training split03Base model04Adapter ΔW05Held-out test06Versioned deployment
The training set changes the model; the held-out set determines whether that change is useful.Scroll the drawing sideways to inspect it.

A useful adaptation experiment starts with a task, a baseline, a dataset, and a reason to believe training will help.

Read the technical perspective /04

Your domain.
The right model strategy.

ACUMEN / ENGINEERING NOTEDifferent paths to model ownershipFIG. HUB
Different paths to model ownershipChoose the smallest change that addresses the actual requirement, then evaluate its operating consequences. Components: Business requirement; Private deployment; Task fine-tuning; Domain pretraining; From-scratch training; Evaluation + ownership. TRAINING CHANGES WEIGHTSTESTING EARNS RELEASE 01Business requirement02Private deployment03Task fine-tuning04Domain pretraining05From-scratch training06Evaluation +ownership
Choose the smallest change that addresses the actual requirement, then evaluate its operating consequences.Scroll the drawing sideways to inspect it.

Private hosting, task adaptation, continued pretraining, and training from scratch have different ownership and resource implications.

Read the technical perspective /05

AI on your terms.
Infrastructure in your control.

ACUMEN / ENGINEERING NOTEA private stack is more than its acceleratorsFIG. HUB
A private stack is more than its acceleratorsApplication access, model serving, retrieval, storage, and monitoring form distinct operating layers. Components: Authenticated apps; Model gateway; GPU serving; CPU / RAM; Vector + document data; Metrics + recovery. SERVING • MEMORY • NETWORK • STORAGE • RECOVERY 01Authenticated apps02Model gateway03GPU serving04CPU / RAM05Vector + documentdata06Metrics + recovery
Application access, model serving, retrieval, storage, and monitoring form distinct operating layers.Scroll the drawing sideways to inspect it.

Memory, concurrency, serving behavior, power, and operational ownership determine whether in-house AI is practical.

Read the technical perspective /06

Useful intelligence.
Deliberate boundaries.

ACUMEN / ENGINEERING NOTEDefense at every trust boundaryFIG. HUB
Defense at every trust boundaryControls surround data, reasoning, and execution. No single model instruction is responsible for all three. Components: Identity + tenant; Data access; Untrusted evidence; Model interaction; Tool policy; Audit + response. NO SINGLE CONTROL OWNS EVERY RISK 01Identity + tenant02Data access03Untrusted evidence04Model interaction05Tool policy06Audit + response
Controls surround data, reasoning, and execution. No single model instruction is responsible for all three.Scroll the drawing sideways to inspect it.

Guardrails are most useful when they reinforce permissions, data boundaries, and tool policies that the model cannot override.

Read the technical perspective /07

Spoken information.
Connected workflows.

ACUMEN / ENGINEERING NOTEFrom sound to a confirmed taskFIG. HUB
From sound to a confirmed taskThe live path includes recognition, turn control, confirmation, and response—not just a transcript endpoint. Components: Audio capture; Speech recognition; Turn + interruption; Confirmed intent; Business tools; Speech / handoff. AUDIO → RECOGNITION → CONFIRMATION → ACTION 01Audio capture02Speech recognition03Turn + interruption04Confirmed intent05Business tools06Speech / handoff
The live path includes recognition, turn control, confirmation, and response—not just a transcript endpoint.Scroll the drawing sideways to inspect it.

Accents, noise, turn-taking, and uncertain numbers shape the workflow behind speech-to-text and voice assistance.

Read the technical perspective /08

Evidence in.
Decisions with structure.

ACUMEN / ENGINEERING NOTEDecisions combine evidence and policyFIG. HUB
Decisions combine evidence and policyModel judgments feed an application-owned rule and threshold layer, with a separate path for review. Components: Evidence state; Defined criteria; Model judgments; Business policy; Permitted outcome; Review / override. EVIDENCEUNCERTAINTY HAS A PATH 01Evidence state02Defined criteria03Model judgments04Business policy05Permitted outcome06Review / override
Model judgments feed an application-owned rule and threshold layer, with a separate path for review.Scroll the drawing sideways to inspect it.

Decision support combines evidence, permitted outcomes, error costs, and a policy for when the system should defer.

Read the technical perspective /09

Beyond a good demo.
Evidence of performance.

ACUMEN / ENGINEERING NOTEA scorecard with separable failure modesFIG. HUB
A scorecard with separable failure modesThese illustrative dimensions are not Acumen performance results. Each layer has its own review criteria. Components: Source / retrieval; Answer support; Tool correctness; Policy adherence; Latency + cost; Human review. SEPARATE CRITERIA • VERSIONED TEST SETS • NO SINGLE MAGIC SCORE 01Source / retrieval02Answer support03Tool correctness04Policy adherence05Latency + cost06Human review
These illustrative dimensions are not Acumen performance results. Each layer has its own review criteria.Scroll the drawing sideways to inspect it.

Evaluate retrieval, model behavior, tool execution, and user outcomes separately enough to locate the failure.

Read the technical perspective
ONE CONNECTED ENGINEERING APPROACH

Business problem.
Whole-system thinking.

Solutions describe the work you want to improve. AI engineering explains the technical depth. Our capabilities connect discovery, software delivery, and the path into production.

Start with a conversation about the problem
A CONVERSATION IS A GOOD START

Let’s put your
ideas to work.

Choose a time to talk, or leave your email and a little context. We’ll take it from there.

Book a 15-minute call
OR LET US GET IN TOUCH
Prefer your email app? business@acumen.llc ↗