7
production systems documented
Live commercial work currently presented on maat.work
Product engineering · Applied AI · Fintech
I build the systems behind the work.
I turn messy operational problems into production software — from financial-advisory infrastructure to local LLM inference and native macOS tools.
Real product surfaces
Not a mockup portfolio.
These are shipped interfaces across construction, membership operations, AI operations and local commerce. Only public-facing or non-sensitive product surfaces appear here; operational records remain protected.
Measured, not implied
A fast credibility check.
Private client systems stay private. I can walk through architecture, trade-offs, tests and failure modes live.
2.1×
faster local inference
13.3 → 27.8 tok/s on a 27B model, output hash-identical
56
tables in a live fintech CRM
Clients, portfolios, positions, risk and compliance
C2
English
Spanish native · Argentina, UTC−3
Selected case studies
Hard problems, shipped all the way through.
01
Financial advisory CRM
A production operating system used by two advisory teams.
Next.jsTypeScriptPostgreSQLPrismaPlaywright
Problem. Replace fragmented spreadsheets and manual workflows without weakening the audit trail required in a regulated business.
What shipped. A 56-table PostgreSQL domain model for clients, portfolios, positions, risk profiles and compliance, with a Next.js product surface and automated test coverage.
02
Apple Silicon LLM inference
More than doubled decode throughput on the same hardware.
MetalMLXPythonQuantizationProfiling
Problem. A 27B model was spending 93% of decode time in one quantized matrix-vector kernel on an M2 Max.
What shipped. Profiled the bottleneck, replaced the kernel in Metal and moved throughput from 13.3 to 27.8 tok/s. Every optimization step preserved hash-identical output.
Benchmarks and implementation available in a technical walkthrough
03
MeetCapture
Private, local meeting capture without a bot or cloud audio.
SwiftSwiftUICore Audiosherpa-onnxSQLite
Problem. Capture system and microphone audio on macOS, transcribe live, diarize speakers and hand off durable summaries without sending recordings to a third party.
What shipped. A native Swift menu-bar app with Core Audio process taps, two swappable ASR engines, live transcription, speaker diarization and atomic handoff.
04
Agent operations platform
A human-readable control plane for software teams and AI agents.
Next.jsReactPostgreSQLRESTMCP
Problem. Coordinate projects, tasks, deploys, prompts and agent activity without hiding operational state inside chat history.
What shipped. An API-first mission control with scoped agent keys, idempotent writes, inbox handoffs, webhooks and a typed MCP surface.
Private production system · live walkthrough available
05
AI creative production pipeline
A local 30-second UGC prototype with a quality gate that can reject it.
MiniMax H3MetalPythonFFmpegWhisperAutomated QA
Problem. Keep identity, scene and audio continuity across a multi-shot generative video while staying inside a 64 GB Apple Silicon compute budget.
What shipped. An end-to-end chain that generates five referenced six-second shots, carries the previous frame forward for continuity, aligns scripted audio, then concatenates, adds word-level captions and renders at 1080×1920. Automated QA scored the current prototype 9/10 visually and still rejected it because voice accuracy failed.
Private production pipeline · rendered outputs and walkthrough available
Public code
Read the work, not the adjectives.
Most client systems are private. These repositories are the shortest path to my engineering decisions.
Software InmobiliariasMulti-tenant real-estate operating system built as a modern TypeScript monorepo.Next.js · tRPC · PrismaMeetCaptureNative macOS local transcription with live ASR and diarization.Swift · Core AudioQuantUcemaFama–French 5-factor portfolio construction and constrained optimization.Python · SciPyQuotaMax RouterOpen-source LLM routing with fallback, quota and budget controls, Prometheus telemetry, alerts and an incident runbook.Python · FastAPI · 81.89% coverage
Complete project atlas
Every repository, organized by the problem it solves.
Public means inspectable. 7 repositories link directly to source.
Private means protected. Commercial systems expose the problem, outcome and stack—not customer data or proprietary code.
54 of 54 projects
Applied AI
MeetCapture
Native macOS meeting capture with local transcription, speaker diarization and durable handoff.
81 local build, smoke, lifecycle, ASR and end-to-end checks passing.
SwiftSwiftUICore AudioSQLite
Active/Public code
Applied AI
HerMaatOS
Local-first operating system for specialized agents, shared work queues, inference routing, voice and a physical interface.
Three-tier local inference gateway; 7,400 requests handled in a measured 24-hour window.
PythonMLXSQLiteMetal
Fintech / Quant
MaatWork CRM
Financial-advisory operating system for clients, portfolios, positions, risk and compliance.
56-table domain model and 700+ test/spec files across the product.
Next.jsTypeScriptPostgreSQLPrisma
Applied AI
Mission Control
API-first operating layer of MaatWork Hub for projects, tasks, deploys, prompts, inboxes and webhooks.
Next.jsPostgreSQLRESTMCP
Product
Construction operations platform
Construction operations platform spanning field reports, certifications, price adjustments and earned-value analysis.
Live system with ARS/USD reporting and AI-assisted document workflows.
Next.jsDrizzleNeonR2
Product
Membership operations platform
Operating system for membership, fees, attendance, cash and electronic invoicing.
Next.jsPrismaNeonARCA
How I work
From ambiguity to evidence.
01
Model the operation
Map the actual decisions, exceptions and data—before choosing the interface.
02
Ship the narrow loop
Put one useful end-to-end workflow in a real user's hands quickly.
03
Measure the constraint
Profile latency, failure modes and adoption instead of optimizing by instinct.
04
Make it operable
Add tests, observability, permissions, handoffs and documentation so the system survives.
For teams building with AI
Make one AI workflow trustworthy before you scale it.
Start with a short, practical checklist. If the workflow is a fit, a scoped AI workflow audit can map bottlenecks, review failure modes and define a measurable improvement; timing depends on the agreed scope.
01
Get the checklist
A compact starting point for retries, evals, permissions and human checkpoints.
02
Share one real loop
Describe the constraint without sending secrets or customer data.
03
Review the fit
A person replies with a bounded scope before any larger commitment.
04
Measure the change
The audit ends with an artifact and before/after evidence—not a promise.
Agent Reliability Checklist
Free · no checkout
Request an audit review.
The checklist is public. Leave an email only if you want a human-reviewed conversation about one workflow.
Open the free checklistFounder price draft
From USD 300
Draft and configurable. Final scope, currency, taxes and terms are confirmed manually. No payment link or checkout is enabled here.
Available for fixed-scope paid work
Need evidence before you need another roadmap?
I take on small, concrete engagements for teams that need a working answer quickly. Every sprint ends with an inspectable artifact, measurements and a clear handoff—not a slide deck about future work.
01
3–5 working days
AI workflow audit
Best for. A team has an AI or agent workflow that is slow, brittle or impossible to trust.
You get. A measured bottleneck map, failure-mode review and one working improvement with before/after evidence.
Evidence: Local LLM inference · agent operations · evaluation and guardrails
02
1–2 weeks
Operational product sprint
Best for. An operation still depends on spreadsheets, repeated handoffs or data no one can safely reconcile.
You get. One narrow production workflow: interface, data model, integrations, permissions, tests and handoff notes.
Evidence: Fintech CRM · construction · membership and billing systems
03
1–2 weeks
Private AI prototype
Best for. A team needs to prove a local, voice, document or creative-AI use case before funding a larger build.
You get. A working prototype on real sample data, with latency/cost constraints, risks and a concrete build-or-stop recommendation.
Evidence: Native transcription · local models · generative-video pipeline
Send the current workflow, the constraint and what a useful result would change. I’ll reply with a bounded scope before either of us commits to a larger project.
Scope a paid sprintEvidence over credentials
No CS degree. No mystery about the work.
I am a licensed financial advisor in Argentina who started writing software because the tools available to advisors were not good enough. That domain knowledge became production software, then a broader engineering practice across AI infrastructure and operational products.
I work by measuring first, making the constraint explicit and proving the result. I use AI heavily in implementation; I do not outsource judgment, architecture or verification to it.
Domain edge
Fintech, advisory operations, quantitative finance
Engineering range
Product, backend, AI systems, native macOS, Metal
The fastest next step