Multi-agent orchestration, multi-model routing, vision pipelines, closed-loop learning, all built on a shared knowledge base the agents and the team both read from. Residential general contractors in production. Not a demo. Not a prototype.
A multi-tenant SaaS platform that covers the full residential GC lifecycle in one place: CRM, estimating, proposals, budgeting, scheduling, daily field logs, change orders, closeout. Built for California GCs at $2M to $20M who escaped spreadsheets or bounced off Procore and Buildertrend. The deeper bet: the data captured during execution becomes intelligence. Estimates get more accurate, schedules learn from actuals, and the bill of materials becomes the input for downstream decisions. The margin you priced is the margin you keep.
Public surface lives at baxiehq.com. The app runs at app.baxiehq.com. In paid beta.
Baxie is the reference implementation. The patterns below are what scoped builds, fractional retainers, and AI-Native Org Audits actually deploy.
A coordinator agent that decomposes a GC's request, routes sub-tasks to specialist agents, and reconciles their outputs against a typed schema. PDF takeoff, estimate generation, schedule synthesis, and scope packaging each have their own agent with their own evals. Failure in one doesn't poison the others.
In consulting: this is the spine of every multi-agent workflow scoped build.
Opus for the hard reasoning calls. Sonnet for the default working tier. Haiku for the routine, high-frequency calls where latency and cost matter more than ceiling. Routing is a config layer, not a hard dependency, so the same workflow swaps models when economics shift.
In consulting: this is the cost-control argument the CFO signs off on.
Architect plans arrive as PDFs. Baxie extracts dimensions, room labels, and assemblies, then feeds them into the estimating agent as structured data. Not OCR on a flattened image. Multi-stage vision plus typed extraction plus validation.
In consulting: this is the pattern for any document-heavy intake workflow, legal redlines, claims processing, lease ingestion.
Field actuals flow back into the next estimate. Schedule slippage becomes a signal that calibrates the next schedule. Change orders feed the assembly cost library. The system gets sharper job by job because every job is a labeled example.
In consulting: this is the differentiator versus bolted-on AI. The loop is the moat.
Most AI fails in production because agents and teams don't share the same context. The shared knowledge base fixes that. Versioned. Searchable. Both your team and your AI read from it. Positioning lives in one file. ICP in one file. Voice rules in one file. Sprint state in one file. When the team changes a decision, the agents inherit it immediately. When an agent generates a draft, it pulls the same canon the team writes against.
Positioning, ICP, voice, and design language are canonical across marketing site, app, and sales. Drift is a bug, not a steady state.
Sprint plans, north-star pillars, and persona rosters live where agents read before drafting anything.
Memory persists between sessions because it lives in markdown, not in a chat thread.
The same shared-knowledge-base pattern Edwin ships in Fractional and Scoped Build engagements.
The growth motion is the proof motion. Architecture posts, worktree screenshots, shipping receipts, and demo clips ship on LinkedIn on a fixed cadence. The receipts compound across three audiences at once: GCs who see the product is alive, operators and founders who want the playbook, and engineers who recognize the architecture.
When Edwin posts about multi-agent orchestration, he posts the diff. When he posts about multi-model routing, he posts the cost graph. The work is the pitch. Marketing voice gets in the way.
Closed-loop systems with humans in them. Augmentation, not replacement. AI-native architecture, not bolted-on. The stakes frame why the build matters. The build proves the stakes are real.
No toy demos. No prompt screenshots without context. The build-in-public surface is constrained to what's already running in app.baxiehq.com against real beta data. The credibility transfers because the patterns are the same ones consulting deploys.
Most AI consultants pitch frameworks. A subset of them ship demos. Almost none of them maintain a production multi-tenant SaaS that exercises the same patterns they're selling. That gap is what Applied AI means: shipped in production, not theorized in a slide. Baxie closes it. The fractional retainer, the scoped build, and the AI-Native Org Audit aren't theoretical exercises. They're the patterns Baxie deploys, transposed onto your stack and your team.
Multi-agent orchestration, multi-model routing, vision pipelines, and closed-loop learning are not slides. They're code paths Edwin maintains in his own codebase. The transfer to a client engagement is mechanical, not aspirational.
Baxie also exercises the org-design questions: where does the human stay in the loop, what work absorbs into the agent, how do evals and observability earn budget. The fractional retainer is the same conversation, scaled to your team.
If a pattern Edwin ships in Baxie doesn't work, the LinkedIn audience sees it that week. The pressure to ship reliable, evaluable, production-grade AI is structural, not promised in a sales call.
baxiehq.com is the live product. The architecture and the receipts are public. If the patterns map to your stack, book a call.