Get the organization ready
I map your systems, data, roles and access, fix the plumbing that would sink an AI rollout, and tell you what to build first and what to skip.
- Systems and data audit
- Access and role model
- Build order and budget
Applied AI · engagements & talks · John Granowitz
Applied AI your people actually use.
I design and build a company's AI layer: custom tools, skills, plugins, agents and models, wired into the systems you already run, with access control from day one. Then I train every seat until it sticks. You get a working system, not a pilot or a slide.
110AI tools in production behind one connector
11 / 11people with access use them every week
1–2 daysfrom a staff request to a working tool
15 → 3days to close the month, after the rebuild
01 · What I build
The whole layer, from data plumbing to the person using it.
Most AI projects fail before the model ever runs. The data isn't trusted, the systems don't connect, nobody has decided who should see what, and nobody trains the people. I do all of it, and I do it myself.
Custom tools inside the assistant they already open
This is a company connector for Claude or ChatGPT. Staff sign in with Google, the connector checks their role, and they see only the tools their seat allows. Those tools read and write your ERP, PM, AP and CAD through APIs and MCP.
- MCP servers and connectors
- Role-scoped skills and plugins
- Documents filled from the source of truth
Multi-step work, with gates
Agents handle multi-step work like month-end close, contract review and addenda, and document assembly. Each one has review and test gates, so a person signs off before anything moves money or leaves the building.
- Agentic workflows
- Human approval points
- Audit trail
Fine-tuning, evaluation, private models
Sometimes a frontier API isn't the answer. I run evaluations before anything rolls out, fine-tune on your own material, and set up local models on your hardware when data can't leave the building.
- Evals first
- Fine-tuning
- Local and private deployment
Training until it sticks
I train in groups and one on one, with skills written for each role. Feedback turns into new tools within days, so people shape what they use. That's why usage keeps growing after launch.
- Group and 1:1 training
- Role playbooks
- Feedback-to-tool loop
Clean the stack underneath
I connect the platforms you already pay for and remove the ones that don't earn their cost. That includes ERP, PM and AP integration, QuickBooks Desktop to Online migrations, and moving AP off paper checks.
- Integrations and write-back
- Migrations
- Subscriptions you can cancel
02 · Case · numbers you can check
Making AI stick in a construction company that didn't ask for it.
I built this from the inside, as head of finance at a commercial subcontractor with $12–15M in revenue. The office of controller, PMs, AP, estimating and procurement ran the company through three systems that didn't agree. It's the same work I now do for other companies.
What I built
First I connected ERP, project management and AP. Then I put 110 role-scoped tools behind one company connector, which staff open from Claude or ChatGPT on a laptop or a phone at the jobsite. The tools fill about 50 branded documents from the source of truth and write back to the platforms instead of re-keying. Each role got its own skills, including contract review, addenda and negotiation. I also moved AP off checks onto Ramp.
Rollout was weekly group training plus one-on-one sessions. Requests became working tools in one to two days, sometimes hours. The senior superintendent was using the tools on site before any field training.
- 11 / 11People with access use it every week.
- ~1 hrSaved per person per week across the office.
- 15 → 3Days to close the month.
- $25k/yrSoftware removed, including a $20k procurement tool and a $5k Procore-to-QuickBooks connector.
- ~10Contract reviews done in-house in the first month, instead of $500 each at counsel.
- 1 seatThe procurement manager left and wasn't replaced. The assistant PM now runs procurement.
03 · How an engagement works
Senior, hands-on, scoped before it starts.
I take a small number of engagements at a time. Each one is fixed-scope and priced before any work begins. You work with me directly, not with an account team.
- ConversationShow me the work that still lives in spreadsheets and the systems that don't agree. I'll tell you plainly whether I'm the right person.
- ProposalWhat I'd build, what I'd replace, what it costs, and what you could cancel. All of it in writing, before anything starts.
- Build on live dataOn your systems, never a demo tenant. Access control goes in first.
- Roll outTrain every seat, group and one-on-one, and turn feedback into new tools as it comes in.
- Hand overYou own the code, the tools and the documentation, and your people can run it without me.
A fit
Owners, CFOs, controllers and operations leaders at companies with real operations, such as construction, field services, finance and professional services. These are companies that already bought the stack and want the AI layer done properly.
Not a fit
If you want a chatbot on the cheap, a staffed agency, a new SaaS login, or "AI strategy" delivered as a deck, I'm the wrong hire.
04 · Speaking & workshops
Talks from someone who built it and rolled it out.
These are practical talks for leadership teams, finance and operations groups, and industry events. They cover what actually worked and what didn't, with real numbers.
- 11 of 11: making AI stick in a company that didn't ask for itAdoption is the hard part. This talk covers how meeting people inside the tools they already use, plus a request-to-tool loop measured in days, got every seat using AI weekly.
- Access control before capabilityRole-scoped AI for real businesses: why sign-in and role checks in front of every tool made it safe to connect the systems that move money.
- From the AR desk to agent fleetsA finance operator's path from QuickBooks and VBA to directing more than 50 coding agents. Domain knowledge matters more than a lab background.
- The agentic month-end closeA hands-on workshop that walks a finance team through multi-step agent workflows, with approval points where they belong.
KeynotePanelHalf-day workshopLeadership briefingRemote or in person
Request a talk
05 · About
I'm John Granowitz. I ran the books, then I built the software.
I've been building with language models since GPT-2 in 2019. That's the same year I started in accounts receivable at a commercial hardscape subcontractor. I worked up through procurement and project management to controller and head of finance, so I know how a company gets paid, buys material and closes its books. The tools I build fit how an office actually works.
Analysync is my practice for that work. It started in 2023 with software integrations for outside clients and is now centered on applied AI. I also build my own products. My construction software, Screed, is built by directing more than 50 coding agents in parallel behind review and test gates. I also run local models on my own GPUs, so I build across the whole range, from frontier APIs to private hardware.
The longer story is at johngranowitz.com.
- Building with LLMssince 2019
- Finance backgroundController · Head of Finance
- StackClaude · ChatGPT · MCP
- SystemsProcore · Ramp · QBO
- ModelsFrontier APIs → local GPUs
- PracticeAnalysync LLC · 2023–
06 · Contact
Tell me what your people still re-key, or what you want your team to hear.
john@analysync.comI read and answer every message myself.