Rashid Azarang

Portrait of Rashid Azarang

Senior AI Engineer

I build AI agents that run in production, the evaluation harnesses that keep them honest, and the data systems underneath them.

My work sits where applied AI meets real operations. I start from how a business actually runs: which decisions repeat, which rules matter, where the data lives, and what has to be true before an automated answer can be trusted. Then I build the agent, the tools it calls, the tests that gate it, and the pipeline that ships it.

Download PDFTechnical dossierCV en español
2010Building software, sites, products, and systems since 2010.
480+ ★GitHub stars across open-source MCP servers, agent tooling, and developer utilities.
GCPAgents in production on Vertex AI, ADK, and Cloud Run, gated in CI/CD.
PaperPublished AI research on context engineering: how agents remember. Open access and reproducible.
Profile

How I work

Agents

Tool-based agents built on versioned prompts, business rules as code, real APIs, and operational data.

Evaluation

Golden sets, deterministic gates in the deploy pipeline, and LLM judges calibrated against human labels.

Cloud

Services that actually deploy: GCP, Cloud Run, Docker, Terraform, CI/CD, IAM, and Secret Manager.

Data

Warehouses, integrations, and MCP surfaces so a team can query its own operation in plain language.

Experience

Selected experience

Jun 2026 - present

Principal AI Engineer, contractor · Laureate International Universities (contract engagement)

Contract engagement on the GenAI program of Laureate International Universities: conversational agents on Google Cloud serving a student population of 700,000+ across two university brands, handling admissions, enrollment, scheduling, routing, and student support.

  • Dominant author of the production agent (150 of 191 commits): prompts, tools, business rules as code, appointment and routing logic, with every change landing as versioned prompt, golden cases, and an ADR.
  • Built and own the evaluation harness — 80+ test modules, 55 golden case files — with online and offline evaluation, calibration tooling, session harvesting, and A/B rigs for context budgets.
  • Calibrated the LLM judge until its verdict could block a release: case-by-case calibration against hand-labeled turns, majority voting, mock-aware temporal anchoring, and a deterministic-only mode for checks a judge should never score.
  • Made evaluation a deploy gate with per-environment depth — smoke on integration, full regression on QA — under trunk-based promotion on Azure DevOps, over a Terraform-managed GCP estate: Vertex AI, Google ADK, Cloud Run, IAM, Secret Manager, Cloud Monitoring.
  • Shipped beyond the agent: the QA bug-tracking app used by testers, and the costs service with recorded parity fixtures so pricing answers stay verifiably correct.
Jan - Jun 2026

Head of AI Engineering, contractor · Talisman, Texas

Contractor for an accounting-software company in Texas, leading applied AI work on the product and shipping a full product module alongside it.

  • Built a project-management module from zero inside the accounting platform: React frontend with feed, board, calendar, gantt, and workload views over an AWS Lambda service on a Neo4j graph.
  • Designed a recurring scheduling engine for month-end close, with cycle resolution and unit-tested date logic, plus cross-company queries and risk scoring for multi-tenant firms.
  • Translated accounting product needs into AI systems, agent workflows, integrations, and technical prototypes.
  • The contract closed when the funding round did not land in time.
2025

Data warehouse and AI integrations, contractor · GreenLight, Dallas

Built the data infrastructure for a property-management company in Dallas.

  • Designed and built a data warehouse consolidating operational information from several systems.
  • Built the integrations that connect those systems to the warehouse and normalize business data.
  • Layered an MCP harness on top so the team could ask questions about its own data in plain language.
2025

Director of Technology · AmFirst, Austin

Technology direction for an insurance company in Austin, Texas.

  • Work across internal systems, data, reporting, automation, and digital operations.
  • Turned insurance processes into flows that could be traced and maintained.
  • Focused on operational continuity, trustworthy information, and internal tools for the team.
2024 - 2026

Own products and AI systems · MetaMCP, Airtable MCP, Crawlio, Mentu

Between contracts I build and ship my own products: SaaS tools, AI systems, MCP infrastructure, and operational software for Mexico, the US, and bilingual markets.

  • MetaMCP: an MCP gateway that collapses many tool surfaces into one endpoint agents can actually consume.
  • Open source: Airtable MCP, ChatGPT Chat Exporter, notion-to-site, and mcp-intelligence, together the most-used code I have published.
  • Crawlio and Mentu: a macOS crawler with a built-in MCP server, and a local-first workflow runtime with reproducible runs and evidence.
  • LeanManufacturing.mx, Cotizera, Dataware, Vaivén, Contratua, Affihub, AyudaLocal, Abastía, and Warming Email in production.
2021 - 2022

Health operations platform · COVID-19 testing

Helped a laboratory sustain high-volume COVID-19 testing during the pandemic by connecting product, data, communication, and daily execution.

  • Supported the scaling of a testing operation under heavy monthly demand while keeping continuity, follow-up, and response capacity intact.
  • Worked on registration, results, user communication, reporting, and operational coordination flows.
  • Production work under real pressure: sensitive data, real users, and continuity that could not slip.
2015 - 2018

Digital business, traffic, and funnels · Early experience

My entry into software came through the internet: sites, campaigns, affiliates, paid traffic, landing pages, conversions, and systems to follow up on prospects.

  • Built digital funnels with landing pages, ads, copy testing, and conversion measurement.
  • Worked in affiliate and performance marketing, where every click, cost, and sale had to be accounted for.
  • Learned to read software as a whole system: acquisition, follow-up, close, operations, and data.
Projects

Selected projects

MCP gateway
MetaMCPGateway that groups and exposes MCP tool surfaces to agents and cloud runtimes.
MCPCloud Run
MCP server
Airtable MCPOpen-source MCP server for Airtable bases, tables, records, and schema.
MCPAPIs
Open source
ChatGPT Chat ExporterBrowser tool that exports ChatGPT conversations to Markdown, JSON, or PDF. The most-used code I have published.
open sourcetools
Consultancy platform
LeanManufacturing.mxSpanish-language lean manufacturing platform: RAG assistant over pgvector, a public MCP server, an OEE calculator, and a five-layer maturity self-assessment.
RAGMCP
Product engineering
Talisman project managementA project-management module built end to end inside a Texas accounting platform, with a recurring scheduling engine for month-end close.
ReactNeo4j
Business workflow
Cotizera AgentsQuote intake, PDF generation, WhatsApp follow-up, and pipeline updates as agent-assisted sales work.
salesevidence
AI crawler
CrawlioNative macOS website downloader and intelligence platform: Swift 6, 25+ actors, 491 source files, with a built-in MCP server for agents.
macOSAI
Media intelligence
VaivénSpanish-language media monitoring for communications and reputation teams.
AImedia
Marketplace MX
AyudaLocalHome-services marketplace built for Mexico, local providers, and WhatsApp.
MXoperations
B2B workspace
ContratuaPrivate workspace for B2B agreements with documents, versions, comments, and traceability.
B2Bdocuments
B2B operations
AbastíaPurchasing and procurement flows for business operations.
procurementops
Data for AI
DatawareManaged warehouse and conversational access over MCP for business data across 50+ integrations.
dataMCP
Data warehouse
Sync Bridge to Data WarehouseSync time cut from 45-60 minutes to 13-20 minutes, with an error rate under 1%.
PostgresETL
Analytics
Enterprise Analytics PlatformMongoDB to PostgreSQL migration, React dashboard, ETL, and analytics for 18,000+ dealers and 50,000+ claims.
BIPostgres
Capabilities

Core capabilities

AI agents

Agents with versioned prompts, tool contracts, business rules as code, functional tests, and evidence for every action taken.

Evaluation and governance

Golden sets, regression gates in CI/CD, calibrated LLM judges, online evaluators, tracing, and the access control around all of it.

Engineering and cloud

Python, TypeScript, REST APIs, Docker, Cloud Run, Terraform, CI/CD, Git, IAM, Secret Manager, and deployments that stay maintainable.

Data and business systems

Postgres, Supabase, warehouses, ETL, connectors, and the operational flows on top: sales, quoting, CRM, reporting — plus the outbound machinery: email deliverability, prospecting engines, and conversion funnels.

Writing

Research and publications

2026

Agent Memory Allocation: Tiers and Policies of Effective Agent Memory · Preprint · Zenodo · DOI 10.5281/zenodo.21938413 · CC BY 4.0

Formalizes agent memory as tiers governed by five retrieval policies and tests the framework with pre-registered, mechanically adjudicated experiments on a production agent system. A public at-power replication (141 documents, 120 frozen questions) found plain file search beat a curated memory index by 12.5 points, with corpus, questions, harness, and adjudicator published for re-running.

2026

Strategic Technical Debt · Working paper

Prices deliberately incurred technical debt as a call option on the validated product, with a demarcation between strategic and toxic debt and a sequential model over belief and debt stock. Theory paper with a specified, not yet executed, empirical program.

2026

Agent Harness: Deterministic Control for Probabilistic Systems · Working paper

Decomposes an agent harness into five planes and proves invariant preservation under complete mediation, once model proposals hold no independent commit authority. Preregistered study, confirmatory result pending. A companion paper measures whether captured knowledge is ever reused in a production agent system.

Ongoing

Technical and product essays · rashidazarang.com

Writing about AI, systems, software architecture, data, operations, tools, and product.

Contact

For roles, projects, or collaboration

Best fit: teams putting AI agents into real operations, teams that need the evaluation and data layer underneath them, or teams building a serious first version of a product.

WebsiteLinkedInGitHub
Schedule a callAgenda