AI agents / product engineering / automation
Perpetual tinkerer and builder. I like understanding why things work, finding where they do not, and making the pieces work better together.
I build AI agents, products, automations, and data systems for messy, real-world work.
Selected work
Working products, private operating systems, and public prototypes. The short version stays visible; architecture, controls, and implementation evidence are inside each technical disclosure.
For a 20+ broker, multi-state team, I replaced a workflow spread across forms, spreadsheets, CRM, Drive, and Slack with a working portal and shared Supabase backend. A reviewed deal or listing now drives financial operations, CRM, content, maps, alerts, and private knowledge.
Open the public demo
Native PDFs and DOCX files are read together. Scanned leases use a bounded OCR fallback; unsupported or conflicting terms stay unresolved.
Executed Lease.pdf + 3 linked documents
Two dates need confirmation. Commencement depends on delivery; expiration cannot be fixed until that condition is satisfied.
| Field | Extracted value | Status | Source note |
|---|---|---|---|
| Tenant | Harbor Dental Group, LLC | 99% / READY | Lease p. 1 |
| Landlord | Westlake Medical Properties, LLC | 99% / READY | Lease p. 1 |
| Premises | 250 Innovation Boulevard | 98% / READY | Lease §1.1 |
| Premises City | Charlotte | 98% / READY | Lease §1.1 |
| Premises State | NC | 98% / READY | Lease §1.1 |
| Premises ZIP | 28203 | 98% / READY | Lease §1.1 |
| Property Type | Medical office | 94% / READY | Lease recitals |
| Suite / Space | Suite 300 | 98% / READY | Lease §1.1 |
| Rentable SF | 18,750 SF | 97% / READY | Lease §1.1 + Exhibit A |
| Commencement Date | Held: delivery condition | CHECK | Lease §3.1; Amendment §1 |
| Expiration Date | Held: 120 months after commencement | CHECK | Lease §3.2 |
| Term (months) | 120 months | 96% / READY | Lease §3.2 |
| Base Rent ($/SF/yr) | $32.00 | 97% / READY | Lease Exhibit B |
| Annual Base Rent | $600,000.00 | 97% / READY | Lease Exhibit B |
| Aggregate Rent | $6,678,318.00 | 95% / READY | Calculated from Exhibit B |
| Rent Schedule | Year 1: $32/SF; annual schedule retained | 95% / READY | Lease Exhibit B |
| Escalations | 3.0% each lease year | 96% / READY | Lease Exhibit B |
| Annual Escalation (%) | 3.0% | 96% / READY | Lease Exhibit B |
| Renewal Options | Two 5-year options at fair market rent | 91% / READY | Lease §18.2 |
| Termination Options | Tenant option after month 84 | 87% / READY | Amendment §4 |
| Notice Requirements | Written notice at least 9 months before term end | 94% / READY | Lease §18.2 |
| Renewal Notice (days) | 270 days | 94% / READY | Normalized from 9 months |
| Security Deposit | $50,000.00 | 96% / READY | Lease §5.1 |
| Free Rent (months) | 4 months | 95% / READY | Amendment §2 |
| TI Allowance (total) | $750,000.00 | 98% / READY | Work Letter §6 |
| TI Allowance ($/SF) | $40.00 | 98% / READY | $750,000 / 18,750 SF |
| Commission Terms | 4.0% of aggregate base rent | 96% / READY | Commission Agreement §2 |
| Commission Rate (%) | 4.0% | 96% / READY | Commission Agreement §2 |
| Gross Commission | $267,132.72 | 95% / READY | 4.0% × aggregate rent |
| Commission Payer | Landlord | 97% / READY | Commission Agreement §3 |
| Broker Splits | Sam Rivera 50% / Morgan Jones 50% | 93% / READY | Commission Agreement §4 |
| Payment Terms | 50% at execution / 50% at commencement | 96% / READY | Commission Agreement §5 |
| Invoice Number | RCB-2026-041 | 99% / READY | Invoice p. 1 |
| Invoice Date | July 22, 2026 | 99% / READY | Invoice p. 1 |
| Payment Due Date | August 21, 2026 | 99% / READY | Invoice p. 1 |
| Invoice Amount | $133,566.36 | 99% / READY | Invoice p. 1 |
More than 635 historical deal records were cleaned and migrated before the old form was retired. Authenticated users now create or review records in Next.js; Zod validation and server actions write canonical deals, listings, commissions, payments, documents, and provenance into Supabase Postgres. FastAPI owns longer integrations and scheduled work.
Closed-deal packets become candidate transaction, commission, and payment rows. For listings, the service reads the Drive folder, OM, photos, and marketing material, then builds the exact required Crexi and CoStar field sets plus Canva flyer inputs. Financial calculations remain deterministic; missing required fields keep creation locked until a person resolves them.
In a separate brokerage deployment, an AI lease intake accepts up to eight executed leases, amendments, commission agreements, deal sheets, and invoices. Native PDFs and DOCX files are read first; scanned pages can fall back to bounded OCR. The draft covers 36 lease, rent, option, notice, allowance, commission, split, and invoice fields with field-level confidence and source context.
Missing values stay null, conditional dates remain conditional, and conflicting documents are shown instead of silently reconciled. A reviewer compares extracted and confirmed values before one idempotent, version-checked database function updates the existing deal, lease, commission, broker splits, field provenance, document links, and reminder rows. It cannot alter the closed stage or date, listing status, or settled payments.
Forms, workbooks, packets, OMs, leases, images.
Zod, parsers, AI extraction, and review flags.
Canonical rows, source artifacts, and provenance.
CRM, content, maps, reminders, crawlers, alerts.
Run status, changelog, review state, and failures.
A persistent Slack agent that investigates questions across the operating database, CRM, Notion, Drive, private documents, and reviewed company memory. It can learn reusable methods and turn a useful question into a scheduled job.
A Python 3.11 Slack Bolt Socket Mode service sends Claude 27 JSON-schema tool contracts spanning Supabase, Follow Up Boss, Notion, Drive, semantic document retrieval, reviewed Slack memory, skills, and saved queries. The service owns the hard eight-call budget, timeout, retry, within-run cache, result caps, and final response.
Reusable methods are upserted to
jarvis_skills by workspace and unique name, with
a description, full body, and one of five validated
categories. Descriptions preload into the prompt; full
bodies load only through skill_read. The
follow-up learning pass skips short questions, short
answers, and zero-tool conversations.
Slack-derived guidance enters retrieval only after a JSON review artifact is approved. Saved questions retain the original prompt, channel, cron, timezone, and optional read-only condition. Database access is protected by a SELECT guard, dedicated read-only Postgres role, transaction mode, statement timeout, and row caps; every run writes one accumulated audit record.
User, channel, workspace, and prior tool context.
Selects bounded tools inside the call budget.
DB, CRM, Notion, Drive, corpus, memory, scheduler.
Tool results and source references return to Slack.
Question, tool trace, latency, outcome, and errors.
A live GRESB advisor backed by an ingested framework knowledge base. The deeper prototype adds tenant-isolated submissions, asset data, benchmarks, evidence, and deterministic Excel scenarios without mixing private firm retrieval into the public corpus.
Direct indicator-code matches and yearly-change guidance are checked before keyword and pgvector semantic retrieval. Public framework and firm-only searches run separately, then merge with source labels and firm evidence first. Empty context produces an explicit gap, never a synthetic answer.
Public queries require firm_id IS NULL.
Firm retrieval calls a dedicated RPC constrained to the
allowed firm ID. The Teams adapter maps a Microsoft tenant
to one firm before retrieval. Unknown IDs fail before the
database search.
EN1 data coverage is floor-area-weighted by ownership share, never asset count. ASHRAE Standard 100:2024 EUI thresholds can deterministically select the Energy Efficiency pathway. LEED O+M changes operational certification coverage under BC1.2. The engine reports the directional benchmark impact and supplied peer average, but reserves an exact point gain because certification scoring also depends on the full peer group and non-public weighting for validation, age, property subtype, and country. Requested scenario files are generated as four-sheet workbooks in code.
Six source-specific MCP servers and one bundled market intelligence interface turn federal economic, demographic, labor, housing, and flood datasets into typed tools an agent can call. Geography resolution, missing-value handling, and live endpoint checks are part of the product, not left to the model.
“Give me a capital-markets snapshot with SOFR, Treasury yields, mortgage rates, and inflation.”
“How are office-using and industrial employment trending in the Dallas–Fort Worth metro?”
“What does the business and tenant base around 4000 Westchase Boulevard look like?”
“For this address, show tract demographics, HUD rent benchmarks, and whether FEMA maps it inside an SFHA.”
The bundled server selects across these source contracts; each source-specific server can also be released and run independently.
FRED exposes six tools, Census/ACS four, FEMA two, HUD six, BLS seven, and County Business Patterns five. A separate CRE market server composes eight high-frequency workflows such as a capital-markets snapshot, metro employment, address-level unemployment, tenant-base analysis, and industry counts.
Shared geocoding resolves coordinates, state, county, tract, and metro identifiers before selecting the matching source tool. Returned payloads retain geography labels and vintage dates so an agent can state exactly what it used.
FRED and ACS sentinels normalize to null. Suppressed CBP employment or payroll is never converted to zero. FEMA distinguishes an unmapped response from low flood risk and converts invalid base-flood elevations to null.
Each source server vendors its HTTP, environment, geocoding, error, and sentinel helpers, so it can leave the monorepo without sibling imports. Live smoke tests check returned values, geography, and payload structure rather than treating any HTTP 200 as success.
A commercial real estate ESG audit engine that turns a property assessment, twelve months of utilities, and basic asset data into calibrated energy, emissions, water, capital, and policy analysis. Every modeled output retains its input and assumption boundary.
Screening outputs, not field observations. Every value below is reproduced in the public methodology case.
Property type, ASHRAE climate zone, construction vintage, and size select a DOE Commercial Prototype Building archetype. CBECS and ENERGY STAR property-type benchmarks validate the starting EUI; twelve months of actual utility data then calibrate the fuel-specific baseline before any conservation measure is evaluated.
Claude extracts systems and utility records from source documents. Python owns measure applicability and DOE prototype deltas, EPA eGRID location-based Scope 2, combustion-based Scope 1, water and DHW linkage, capital cost, incentives, payback, IRR, NPV, and building performance standard exposure.
Every output is labeled observed, calibrated, modeled, or unavailable. The public case uses synthetic property inputs. Jurisdiction exposure appears only when the policy table and asset applicability are current; unsupported costs, incentives, and savings stay withheld instead of becoming polished but ungrounded certainty.
An open-source local MCP server that connects Claude directly to the official EPA ENERGY STAR Portfolio Manager API. It discovers properties and metric names, normalizes XML into predictable JSON, and stores nothing.
The stdio MCP process reads local environment variables and sends credentials only to the official ESPM endpoint. There is no application backend, telemetry service, or project database.
Portfolio Manager exposes XML and metric names that can change by report. The server discovers available metrics, requests the correct reporting year, then normalizes results into stable JSON contracts for the model.
The repository documents working test and live account paths but does not currently ship an automated test suite. Public portfolio responses demonstrate the interface, not a claim about a real user's property count.
Arthur is my VPS-hosted technical PM. It triages client requests and operating signals, creates evidence-backed tickets in one canonical Work Queue, prepares implementation briefs for the right coding agent, and reconciles completion against GitHub, checks, and the live system.
RCB email thread · named requester · desired outcome · source reference · client and urgency resolved
Stable source key · Notion Work Queue · owner: Codex · next action: implementation brief · live verification required
GitHub and production evidence update the ticket; the linked private Slack thread confirms the final state
Scheduled sweeps inspect explicit KWC Slack requests, RCB email requests, Granola commitments, unfinished Codex or Claude handoffs, GitHub evidence, and verified failures. Arthur requires a requester, outcome, source, stable key, owner, next physical action, and due-date state before creating or updating a ticket.
Arthur identifies the likely repository, preserves the evidence and constraints, drafts the proposed approach, acceptance checks, and live-verification requirements, then recommends a Codex or Claude Code handoff. Coding, merging, and deployment remain separate authorizations.
Notion is the canonical Work Queue; SQLite holds only cursors, locks, dedupe, and runtime state. Each surfaced ticket gets a private Slack card and thread. Replies such as start, waiting, done, or a correction update only that linked ticket and confirm the resulting state in-thread.
Code tickets close only after a merged PR, passing checks, and live evidence when required. Arthur sends one bounded, changed-items-only private brief at 8:30 AM ET and immediate exceptions only for verified failures or urgent requests. Client messages, merges, deploys, invoices, and calendar writes remain approval-gated.
Supporting systems
The portfolio extends beyond the flagships: private document retrieval, document operations, and a multi-tenant lead system.
Push notifications, a modified-since backstop, bounded OCR, batched embeddings, filtered retrieval, and citations back to the source file.
DRIVE WATCH / OCR / VOYAGE / PGVECTORRent-roll and P&L extraction with reconciliation and review flags, plus format-locked generation that produces byte-identical output from the same structured input.
PDF / XLSX / JSON / DETERMINISTIC QCA multi-tenant qualification and estimate-booking system with provider adapters, contractor-scoped consent state, image interpretation, and real Google Calendar availability.
FASTAPI / POSTGRES / CLAUDE / TELNYX / CALENDARThat same instinct carries beyond software. Chances are you can find me adding to my garden, renovating part of my home, or tinkering with a project car or boat.
Get in touch
Want to talk through a project, compare notes, or just connect? Shoot me an email.