Skip to content

Latest commit

 

History

History
585 lines (339 loc) · 21.6 KB

File metadata and controls

585 lines (339 loc) · 21.6 KB
theme default
highlighter shiki
fonts
sans mono weights
DM Sans
JetBrains Mono
400,600,700
title Creating Robust Agentic AI Systems to Increase Data Insight Accessibility
info **Creating Robust Agentic AI Systems to Increase Data Insight Accessibility** *Also known as: Building Agentic Solutions for Construction* Analytics leaders are pushing **data-driven decisions** across the enterprise—for AEC that still means **field, trailer, and finance** working off the same facts. **Accessibility of insight** (not only raw tables) is the bottleneck; **agents and generative surfaces** help when outputs stay **accurate, grounded, and governable**. Robust agentic systems and accessible data insight for AEC.
class text-left
transition fade-out
mdc true
duration 120min
Part 1 · Intro
Construction · Agentic AI · Data insight

Creating Robust Agentic AI Systems to Increase Data Insight Accessibility

Also known as: Building Agentic Solutions for Construction

Accessible insight · Analytics adoption · Production-grade agents

Part 1 · Practitioner

The practitioner’s journey

HPC foundation (~10 years): shipping software products, taking them to market, and learning hyperscale reliability—customers across manufacturing, life sciences, oil & gas, aerospace, defense, and automotive. That shaped how I think about scalable systems.

Cloud shift: as HPC workloads moved cloudward, the work became architecting cloud-native applications and native integrations—the same discipline, different substrate.

FlyPaper (CTO): BIM coordination, superintendent dailies, and pull planning software grounded in the Last Planner Systemmobile and web products built to meet crews in the field.

Procore acquisition → Barton Malow: helped stand up data and automation, built the engineering team, ran early “what is AI / how do we use it?” research, shipped internal orchestration work, and moved with the industry toward MCP-style tool contracts.

Today: building agentic AI solutions for AEC strategy professionals—with a special focus on competitive intelligence tools that empower BD and strategic decision-makers, as well as on enhancing preconstruction workflows for greater efficiency and insight.


Part 2 · Strategy 0:10–0:45

Orchestration — the central nervous system

In 2026 the battleground is who controls context and workflow—not which foundation model “wins.”

Strategy: choose off-the-shelf when the problem already fits the vendor’s graph and speed beats differentiation. Choose hands-on when workflow, data, economics, or deep internal business knowledge must be yours—or you share the same packaged intelligence as everyone else.

LLM = engine · orchestration = transmission

Off-the-shelf

  • Databricks Agent Bricks + Unity Catalog — agents, tools, and lakehouse data in one governed loop (identity, MCP catalog, observability)
  • Procore Helix (Assist, Agent Builder, Datagrid / connectors)—excellent in-platform loop; weak where edge lives in local LOB spreadsheets / systems outside the graph
  • Glean, Bedrock Agents — “managed factory”: you land data; they run the thinking loop (COIs, standard retrieval)
  • Autodesk Assistant, Claude Cowork, Zapier-class hubs — utility productivity; vendor tax / metering for speed—fine when the vendor boundary is acceptable

Hands-on (power move)

  • Workflow you encode — treat logic as a state machine, not a wall of prompt. Example: submittal — spec → extract → compare → if deviation > threshold → human. Stack: LangGraph / LangChain, AWS Agent Core, Temporal + n8n/Windmill for durable glue
  • Typed + multi-agentPydantic AI (ERP-safe I/O), CrewAI (roles and handoffs)
  • Integration boundarycustom MCP (Python/Node) for wholesale APS / Procore APIs vs throttled vendor MCP; connect internal + third-party systems and lake/warehouse views so stakeholders can ask data without living in SQL—full story on the MCP slide
  • Product surfaceVercel AI SDK + bespoke UI (e.g. superintendent: Check safety · Summarize daily · Order materials)—tools, not a thread. Heatmap / silent auditor patterns — next slide
  • Capacity under load — e.g. Bedrock provisioned throughput when crunch beats public rate limits (inference slide)
TierStrategyAdvantage
Off-the-shelfBuy the featureSpeed; low kit; vendor-shaped workflow
Hands-onOwn workflow + stackMargin; moat; full context boundary
Own the policy graph and context boundary—web, MCP, API, SDK are doors into the same graph Orchestration

Part 2 · Strategy

Beyond the chatbot

Role-appropriate surfaces — not “chat bad,” but “chat is one option”

Conversational analytics: when the surface is chat, context and specificity (what project, what time range, which source of truth) directly shape output quality—brief users once; don’t assume “the AI knows.”

Field vs. trailer: voice-forward chat can fit hands-busy field roles; a project engineer on an iPad or laptop in the trailer often needs structured views, approvals, and diff—not a thread.

When chat is the wrong default: if the job is “produce an artifact” (report, package, leveled bid), ship the artifact UI—a dashboard or wizard that thinks, instead of making people type prompts.

Labor study — product vs. chat

Not only a chatbot that “explains BLS.” Ship an interactive labor study—BLS (and peers) fused with your assumptions, AI synthesis, exportable narrative—so users manipulate the report, not the prompt box.

Submittals — distribution lists vs. experience

Classic email + distribution lists work until volume and revision churn win. A unified web/mobile submittal surface can abstract AI behind routing, redlines, and audit trails—outcome-first.

Example — submittal reviewer: drag a PDF; get side-by-side redline and citations—not a chat transcript.

Philosophy

Make AI invisible: enrich the artifact people already owe—RFIs, packages, studies—not “another app to log into.”

Deliver formats teams already live in—not a universal “learn to prompt” tax.


Part 2 · Strategy

Vendor agents vs. rolling your own

MCP is one way to wire tools — the business question is who owns the boundary

Vendor path: Procore / Autodesk / others ship agents, connectors, and “opinionated” flows—fast to adopt, often metered, roadmap is theirs. You pay for the abstraction; it may not optimize for your margin.

Build path: Custom integrations (including MCP servers) across internal systems and third-party platformsAPS, ERP read models, lake / warehouse views—your auth, caching, pagination, and policy. Sometimes you can beat packaged assistants on latency and unit economics—sometimes not; prototype both.

Decision lens: technical capability and strategic fit; integration surface area (how many systems?), compliance, rate limits, and whether the vendor’s “happy path” matches your WBS / CM reality.

Own the boundary you cannot afford to rent: lineage, approvals, and where context is assembled.

Pick vendor agents when the fit is real; build when the fit is strategic or the bill/risk is structural Integration

Part 2 · Strategy

Inference, tokens, and scale

Where you run matters as much as which model

flowchart LR
  subgraph paths [Call paths]
    D[Direct vendor API]
    H[Hyperscaler hosted Anthropic OpenAI Google]
  end
  D --> T[Tokens billed per call]
  H --> T
  T --> K[Demand spikes caching regional residency]
Loading

Prototype vs. production: direct Anthropic / OpenAI APIs for benching; Bedrock / Azure OpenAI / Vertex for procurement, guardrails, and data residency.

Economics move: Anthropic models run on AWS—list vs. Bedrock pricing crosses over over time as deals change. Re-benchmark quarterly; the “cheaper path” is not a constant.

Tokens = capacity planning: batch size, concurrency, prompt caching for stable corpora (spec books), and back-pressure when vendors throttle.

Security story: keep payloads in your boundary; enterprise agreements for training use—give IT one crisp sentence.


Part 2 · Strategy

Model selection — right engine, right hill

Models are stochastic—reliable systems come from router / worker / judge, schemas, and eval loops, not from pretending the LLM is a calculator.

flowchart TB
  R[Router small 8B-class] --> W[Worker 70B-class or API]
  W --> J[Judge frontier when risk is legal or contractual]
  W --> L[Local or HF endpoint for repetitive extract]
Loading

Use the smallest reliable model per step — tokens are margin, and latency is trust.

Hugging Face & self-host: great for exploration, air-gapped constraints, and high-volume narrow tasks—know your ops cost, not just API list price.

Anti-pattern: “Claude for everything” is using a haul truck to catch a mouse—fun demos, expensive production.


Part 2 · Strategy

The fine-tuning trap

flowchart TB
  A[Instruction-tuned prompts 5 to 10 gold examples] --> B[Max context plus caching]
  B --> C{Need FT?}
  C -->|Usually no| D[Ship evaluate iterate]
  C -->|Sometimes| E[Strict JSON contract or 1M plus tiny tasks daily]
Loading

Before fine-tuning, exhaust: ICL + full context (manuals, standards packs) with caching where available.

Fine-tune when: you need a hard format contract or massive repetitive micro-tasks where latency/cost dominates.

Trust before weights: guardrails, structured outputs, and eval + reasoning loops with human gates—instruction and ICL usually move accuracy more than fine-tune; don’t erode stakeholder confidence chasing novelty.

Motto: context windows first; weights are a secondary lever.


layout: default class: tp-scenario-slide tp-scenario-slide--stacked

Part 3 · Workshop

Scenario cards

Group 1 — Silent submittal auditor

Problem Reviewers miss alternates buried in long vendor PDFs.

Task Agentic flow that flags mismatches against the master specwithout a chat-first UI.

Group 2 — Bid leveling engine

Problem Twenty trade bids with different exclusions and scopes.

Task Use MCP to pull APS model or property data and normalize bids into a standard WBS view.

Group 3 — Historical knowledge estimator

Problem Junior staff lack narrative memory of why a past job bled margin.

Task Tool that reads actuals from ERP via MCP to sanity-check a new estimate.

More workshop ideas — swap in for a future session or spare group

Alternate — Invoice vs. contract

Problem Draws, pay apps, and vendor invoices drift from contract / SOV language—finance catches it late.

Task Agentic line-item reconciliation (invoice ↔ contract ↔ change orders) with an exceptions queue and citations—not a chat thread.

Alternate — Data cleanup synthesis

Problem Exports are noisy: duplicate vendors, bad codes, inconsistent units—BI is downstream of garbage.

Task AI-assisted normalization (suggest mappings, cluster dupes) with human approval gates; ship clean dimensions to the warehouse.

Alternate — AEC news → strategy brief

Problem BD and leadership drown in feeds; strategy memos go stale before anyone reads them.

Task Synthesize AEC news (plus filings, press, selected trade pubs) into a strategy document template—markets, competitors, risks—with citations and refresh hooks.


Close · Debrief

Critique rubric

After groups present — use this to score designs in ~2 minutes each. Favor ideas with a clear pilot metric (time saved, defect rate, adoption) so spend stays justified.

  1. Architecture: router/worker/judge — or one monolithic call?
  1. UX: chatbot by default, or a tailored surface?
  1. Data: what is the source of truth and who can access it?
  1. Monday plan: what is the 30-day pilot (scope, metrics, owner)?
  1. Outcome: what measurable result unlocks the next funding or scale gate?

layout: center

Final thought

Your job is not to be an AI user. It is to be a practitioner: build bridges, own context, and protect your company’s intelligence.

Data insight accessibility · AEC · Own the graph

https://www.linkedin.com/in/ispyhumanfly/ or email me at dan@thoughtpivot.com.