HECTOR ARZATE

A construction cost-estimating product, born as a C++ desktop app in 2009. I lead its rewrite into a modular web suite — budgets, planning, purchasing, estimates — that construction firms and government agencies use to bid and run public works.

My role Product + systems lead: I design each module end to end, build the first working version, and hand it to the dev team, who take it to production. I review, integrate their branches and keep CI green.

• Mapped the legacy C++ app screen by screen into web modules: budgets, planning, purchasing, estimates
• .NET 10 / ABP + React modules, Node portals, one sign-on across the suite
• Desktop customers keep working while modules move to the web
• Runs in four languages (es / en / pt-BR / fr)

Three AI apps around the bid: one finds public tenders and assembles the proposal file, one reads the tender documents and drafts the cost estimate, one turns construction drawings (DWG/DXF/IFC) into quantities and a budget.

My role Sat with estimators to see how a bid really gets made, designed the flow, prototyped each app working end to end, then handed them to the team to harden and ship.

• Reads tender PDFs and scans; extracts the catalogue of work items
• RAG over the company cost database (Postgres + pgvector) to price each item
• Drawings → quantities → budget
• One credit wallet shared across the apps; usage and real AI cost tracked per client

The control center behind every app in the suite: licences, self-service sales and renewals wired to Odoo (order → invoice/CFDI → payment, with no human in the loop), distributors, teams, and the AI gateway every app goes through.

My role Designed the hub and the sales automation with the people who run billing and support, mapped Odoo’s licensing rules, built and tested the flows end to end before they went to production.

• Customer buys or renews in the portal; Odoo invoices it and the licence activates automatically
• Distributor portal; institutional clients upload their paperwork, read by a local model
• AI gateway: model routing, credits, live cost per client and per model
• MCP catalogue so assistants can drive the whole suite through one audited door

An AI terminal inside the product that runs every module from plain-language commands on a small local model, plus a public sales assistant that answers questions and lets visitors buy from the chat.

My role Defined what “good” means with support and sales, built the eval and the dataset rules (anonymize first), trained and measured each version, and wired it into the product behind flags.

• Evaluated local models on 170+ real tasks; scored +1 right / 0 asks / −1 invents
• LoRA fine-tune on ~4,000 examples from manuals, academy videos and anonymized support tickets
• Guards: no invented ids, confirmation before anything writes or spends
• Sales assistant hands off to the existing checkout (OTP + Stripe)

The reference cost database estimators price against — tens of thousands of materials, labour and equipment costs — rebuilt so it updates itself for each country: a pipeline plus AI agents that research and verify local prices, from Mexico to 11 more countries.

My role Designed the data model, the quality gate and the agent workflow; built the pipeline and supervised the agents country by country.

• Pipeline: dump → SQLite → per-country seed → patches → quality gate
• Recalculation engine matches the Mexican original at 99.97%
• A country passes only with real, dated, sourced prices and local units and labour costs
• Teams of AI agents (leads + workers) research each country in parallel

Multi-tenant SaaS platform for coffee roasteries and connected specialty coffee shops — inventory, production, B2B sales, contract roasting, and automated reordering between roastery headquarters and retail locations.

My role My own product: I found the problem inside a coffee company I’m acquiring, designed the whole operating system and built it myself.

• Roastery: green coffee inventory, contracts, production scheduling, packaging, stock control, B2B orders
• Coffee shop: point of sale, menu management, sales metrics, automatic reorder from HQ
• Next.js 14 App Router, Drizzle ORM, PostgreSQL with Row-Level Security
• Local admin app built with Tauri
• Per-tenant subdomains with role-based access — owners, managers, staff, baristas
• Contract roasting workflows and automated replenishment across roastery ↔ retail

A game starring the BEARSHARK Company cast — the same characters sculpted, retopologized and rigged in Blender for the animation, now being brought into a playable world (a ship set is next).

• Characters built once in Blender, shared by the animation and the game
• Rigged and animated models exported to glTF for real-time engines
• In production: ship set, levels and gameplay come next

A rewrite of an open-source drawing and painting app: a Photoshop-style layered editor for painters, ported from Swift (macOS only) to C# so it runs on Windows and Linux, with a file format for canvases bigger than the original allows.

• C# / .NET 10 + Avalonia + SkiaSharp, ported from the Swift original (MIT)
• Layers, folders, masks, clipping and editable adjustment layers
• Pressure brushes and presets; ONNX background remover
• Saves .comp (opens in the Mac app) or .compx for huge canvases

A desktop app for storyboard artists: drop in boards, time each drawing, sketch notes on layers, sync to a dialogue track and export the animatic as video.

• Electron + React + TypeScript + Vite
• Boards on a timeline, held for N frames each
• Notes on layers with pressure brushes and onion skin
• Dialogue track with waveform; export MP4, MOV or PNG sequence via FFmpeg

A Litecoin-style cryptocurrency fork built entirely out of boredom — scrypt proof-of-work, one-minute blocks, and 69 coins per block.

• Bitcoin/Litecoin-derived codebase with scrypt PoW
• ~690 million total supply; subsidy halves every ~4 years
• MIT licensed — integration and staging tree on GitHub

GitHub ↗

A written paper on molecular time travel, a Python proof-of-concept, and a plain-language explanation of how the idea could work in practice.

• Full paper (PDF) outlining the molecular time-travel hypothesis
• Test01_MTT — Python scripts that exercise the core idea
• Walkthrough of the mechanics and what a real-world version would need

GitHub ↗Paper (PDF) ↗

CryptorPunks (CHP) — a live Ethereum minting dApp: React + Web3 frontend wired to an ERC-721 contract for a 1,000-piece collection at 0.0069 ETH, deployed and listed on OpenSea.

• React/Redux dApp with MetaMask connect, supply/cost state, and on-chain mint
• Ethereum Mainnet contract (CHP) — max supply 1,000, public mint at 0.0069 ETH
• Static production build hosted with config-driven ABI, theme, and OpenSea collection link

GitHub ↗Etherscan ↗OpenSea ↗