i.
Custom Extractors
Trained on your documents — invoices, sponsorship contracts, insurance forms, freight bills. Schema-bound LLM output with deterministic post-validation.
№ 002
The Practice — disciplines, methods, deliverables
Four disciplines.
One workshop.
Each engagement is led by a senior practitioner — no handoffs, no juniors learning on your problem. The disciplines below frequently overlap; most projects pull from at least two.
§ 01 Discipline I
The Headline
Agentic AI Systems
Agentic AI is not a chatbot with extra steps. The systems we build reason over goals, call tools against your real infrastructure, and write back to ground-truth ledgers — with audit trails the finance team can defend.
Architecture Posture
Hybrid agentic + deterministic — by default.
Pure-agent systems hallucinate themselves into expensive corners. Pure-deterministic systems can't handle the messy 20% of any real workflow. We design around that boundary on purpose: an LLM-driven planner orchestrates deterministic tools, with human-in-the-loop gates at the steps that matter.
Stack
- FastAPI & Pydantic
- Google ADK · Gemini
- Cloud Run · Cloud Tasks
- Anthropic API · MCP
- Postgres · Firestore
Patterns
- Multi-agent orchestration
- Tool-calling & schema enforcement
- Async journal pipelines
- Retry & idempotency
- Observability & replay
§ 02 Discipline II
The Backbone
Document Intelligence
Most enterprise data hides inside PDFs, scanned contracts, and email attachments. We build the extraction, retrieval, and verification pipelines that turn that surface area into structured, queryable, postable records.
ii.
Retrieval (RAG)
Private-corpus retrieval over policies, manuals, contracts. Citations that point back to the source, every time. No confident-sounding fabrications.
iii.
System Integration
QuickBooks Online, Salesforce, Workday, custom ERPs. Extracted data flows back into the system of record — not into another dashboard nobody opens.
§ 03 Discipline III
The Foundation
Full-Stack Platforms
AI features are nothing without a platform that boots, scales, and stays observable. We build the React/Node/Mongo/GCP foundation — and have the deployment scars to prove it.
Reference engagement · Live, present-day · Codename
Project Citadel — events at the upper register
A platform serving heads-of-state gatherings and the upper register of professional convenings. Master-detail and accordion UIs, Twilio SMS verification for door check-in, calendar-aware RSVP (Google / Outlook / Apple), MongoDB Atlas index work that identified missing indexes as the root cause of scale issues, and App Engine observability for live deployments.
§ 04 Discipline IV
Off-Ramp
Advisory & Architecture Review
Sometimes the right answer isn't another sprint — it's a senior review before you commit to one. Engagements at this tier cover AI strategy, system architecture, security posture, and team structure for groups already shipping.
Format
Two-week deep dive. Read the code, talk to the team, ship a written report.
Deliverable
Architecture review, prioritized risk register, sequencing recommendation. Yours to keep.
Posture
Plain-spoken. Trade-offs named. No vendor pressure — we're not reselling anyone's platform.
§ 05 How We Work
—The engagement—
Step i.
The Brief
A 60-minute call. We listen, ask the questions that matter, and tell you whether we're the right shop — or who we'd recommend if we're not.
Step ii.
The Proposal
A short document — scope, milestones, fee, what's explicitly out. Reviewed and signed within a week.
Step iii.
The Build
Code lands in your repo on day one. Weekly demos. Async-first updates. No surprises at the end.
Step iv.
The Handoff
Written documentation, runbooks, and a transition window. We stay on retainer as long as you want, and step away cleanly when you don't.
§ § §