# marcusnavarrogabrich.com

> llms.txt — summary of this site for language models.

## Marcus N. Gabrich

Systems Architect, Tech Lead, and Computer Scientist (BSc CS, PUC Minas 2019–2025, GPA 84.16;
EPITA Paris exchange Feb–Aug 2023, semester average 15.52/20). Tech Lead, Principal Systems
Architect & General Manager (Interim CEO) at RoiGenius — a Revenue Attribution SaaS running
8 NestJS 11 microservices on Google Cloud Run + Cloud SQL PostgreSQL 15, ingesting webhooks
from 89 ad/checkout platforms with string-based BRL precision.

Independent technology consultant for Assiny (checkout & payments). Quantitative systems
developer (MetaTrader 5, Python/MQL5).

Research & awards:
- SSCAD 2024 (published 23 Oct 2024) — "Neural Network Analysis for CRISPR: A Quantum Computing
  Approach"; hybrid quantum-classical architectures (Qiskit, PennyLane, IBM Quantum).
- ABI AcademyHack Ambev, Nov 2021 — "Discount Leverage", 3rd place overall (Recommendation &
  Optimization Systems category).

## S.A.C.I. — Sistema Automático de Caminhos Inteligentes

S.A.C.I. ("Automatic System of Intelligent Paths") is the decision layer described at /projects/calibrated-decision-layers/saci/: it
takes a program's structured state plus typed questions with bounded answer spaces and returns
typed answers with calibrated probabilities and a confidence band (auto / human / defer).
Application code acts on the band, never on a bare argmax, so abstention is a first-class outcome.

The design borrows Kahneman's System 1 latency contract (single-pass, sub-50 ms on the async
path), a graph/knot metaphor for state-to-option routing, and adaptive redirection — including
routing around a decision rather than forcing it. State, routing and execution are deliberately
three separate layers; the model never owns a consequence.

S.A.C.I. is an independent implementation of the typed-decision pattern, built and measured here.
It is not a wrapper around another system, and no third-party system is part of its architecture.
Other implementations of the same pattern are treated as comparison points:
/projects/calibrated-decision-layers/saci/benchmarks carries a table comparing our measured results
against four named public reference systems on the public JevBench listing (v1.3.0), with reference
numbers (read 2026-09-21) kept structurally separate from our own measurements, plus a stated
reading of where we are not first. The hard split does not converge: the listed systems have a
median of 0.491, encoder-class systems span 0.318–0.500, and the best listed result is 0.950. An
earlier reading ("everyone converges around 0.30–0.38") was wrong; it is retracted publicly as S9
in /projects/calibrated-decision-layers/saci/log.

Reported numbers (held-out splits, Expected Calibration Error, lower is better): best-calibrated
question ECE 0.019 at 90.3% accuracy; a second question at ECE 0.047 / 90.4% with 3.48x
abstention enrichment at 80% coverage; a trading-action question that missed the pre-registered
ECE bar (0.3427 vs 0.08) and is published as rejected. On the public benchmark sample, the encoder
baseline scored 0.443/0.300 (easy/hard) and improved to 0.667/0.350 with statement framing — a
22-point gain at no accuracy cost elsewhere. Comparison rows are served from per-question id
intersections.

The page also documents the benchmark methodology: time-blocked splits, rolling-origin
walk-forward, leakage audits, Wilson lower bounds instead of point estimates, split-half
re-measurement to defeat the winner's curse, trials-adjusted deflation, and asymmetric-cost risk
control.

The project is in training: the reinforcement stage, a rebalanced retrain and the domain-specific
public benchmark are open items, listed on the page rather than omitted.

## Projects

Each project has its own case study at /projects/<slug> — context, problem, approach including what
did not work, measured outcome, and what the author would do differently:

- /projects/roigenius — revenue attribution SaaS: eight microservices in production, 89 integrated
  advertising and checkout platforms, fixed-point money handling, asynchronous webhook ingestion
  with dead-letter queues.
- /projects/local-ai-platform — a self-hosted, offline-capable platform for model serving, routing,
  bi-temporal memory and agent governance.
- /projects/discount-leverage — discount personalisation as a multi-armed bandit problem;
  3rd place overall at ABI AcademyHack Ambev (Nov 2021), public repository.
- /projects/crispr-quantum — published paper on hybrid quantum-classical neural networks for
  CRISPR data (SSCAD 2024, published 23 October 2024); method at /method/, full results table at /results/.
- /projects/trading-systems — an automated gold-trading program on a demo account (DEMO; no money
  results published), with method and change log; unattended execution with independent risk
  containment.
- /projects/calibrated-decision-layers — the decision layers, one folder per component:
  - saci/ — S.A.C.I., the route generator: naming, architecture, roadmap, open items
    - saci/method — questions, the contract, decision paths
    - saci/benchmarks — fair-comparison rules, methodology, the public benchmark vs published results
    - saci/log — every shipped change, with public corrections kept in place
  - saco/ — S.A.C.O., the context filter: which transcript segments an agent keeps
    - saco/results — held-out results, the promotion rule, future sessions, exit criteria
    - saco/benchmarks — latency, footprint and training cost against fixed ceilings
  - injection-screen/ — the check on every tool result, against its latency ceiling

## Benchmark methodology

/projects/local-ai-platform/benchmarks/ documents how models are chosen here: a 57-task, 12-domain suite with three scopes, and
five verification methods ordered by how much they can deceive — deterministic matching, structured
parsing, sandboxed execution of generated code, rubric judging by a separate model, and measured
efficiency kept as its own axis. Anti-contamination uses per-run dynamic needles, verbatim official
items where licences permit, fixed sampling parameters and an append-only ledger.

A/B interventions against the context pipeline follow a pre-registered protocol: the task list is
fixed before the run, arms alternate, and the kill rule is "if the billed total goes up, the
intervention gets turned off", with a keyword correctness floor so cheaper cannot mean worse.

Capacity is measured per configuration with a resolvable evidence pointer; verdicts include
`do_not_use`, which makes the preflight refuse to start a run — measured knowledge blocks work
instead of describing it. Energy per token is measured on the same run as throughput.

Reported limitations include: no real repository-level agentic benchmark, environment isolation is
not network isolation, a local judge is still a model, and none of these scores are comparable to
vendor full-precision results.

### Benchmark methodology

How a number here earns the right to be believed: time-blocked splits (never random on temporal
data) with rolling-origin walk-forward and a leakage audit as an explicit step; ECE over 10 bins as
the primary gate, with abstention reported as coverage and error-enrichment together and AUROC
alongside accuracy; latency measured separately, never traded against quality.

Gates are pre-registered before the run, use Wilson lower bounds rather than point estimates, and
apply two corrections that matter: split-half re-measurement (thresholds fit on one half, reported
on the other, 200 repetitions — the step that once turned a previously adopted rule into a negative
round) and trials-adjusted deflation, where every prompt/threshold/seed variant counts as an
attempt. Where the two error directions cost differently, the threshold bounds a weighted risk
instead.

Comparisons intersect items by id before ranking; third-party published numbers are labelled as
reference and never mixed with measured rows; every negative round records the artifact that
invalidated it. The public artifact (dataset card, harness, redacted dataset, seeds) is pending —
the practices already run locally.

## Publications

- "Análise de Redes Neurais para CRISPR: Uma Abordagem com Computação Quântica" (Neural Network
  Analysis for CRISPR: A Quantum Computing Approach) — 25th Symposium on High Performance
  Computing Systems (SSCAD 2024), pages 13-24, published 23 October 2024, in Portuguese. Predicts
  CRISPR gene dependency from gene copy number (DepMap) with a classical and two hybrid
  quantum-classical networks, quantum circuits simulated on classical hardware. All three R2 values
  are negative; the hybrids reach higher accuracy at a far higher simulation cost.
  Indexed on ResearchGate. Case study at /projects/crispr-quantum/.

## Research — calibrated decision layers

Work on the class of models that answer typed questions with calibrated probabilities instead of
generating prose: bounded answer spaces, expected calibration error (ECE) as the gate, conformal
thresholds per question so abstention is a first-class outcome, and application code owning the
combination of answers. Benchmarks are published with method, sample context and verdicts against
a pre-registered bar (ECE ≤ 0.08) — including a configuration that failed it. See /research.

## Pages

- / — Marcus N. Gabrich — Systems Architect · Tech Lead · Computer Scientist. Revenue-grade microservices and local-first AI infrastructure.
- /activity/ — Public engineering activity: contribution history, decision-layer benchmarks and aggregate pipeline telemetry.
- /colophon/ — How this site is built, which gates block a release, the performance budget, the correction policy and what is disclosed about AI assistance.
- /cv/ — CV — Marcus N. Gabrich: systems architecture and tech leadership at RoiGenius, published research (SSCAD 2024), BSc Computer Science at PUC Minas.
- /now/ — What I'm working on right now: hardening the revenue attribution platform, SFT experiments on self-hosted compute, and harness engineering for autonomous agents.
- /projects/ — Selected work: revenue attribution in production, a local-first AI platform, a decision layer benchmarked in public, and a published HPC paper.
- /projects/calibrated-decision-layers/ — Calibrated decision layers: S.A.C.I. routing, the S.A.C.O. context filter and an injection screen — components that answer with calibrated probabilities, not prose.
- /projects/calibrated-decision-layers/injection-screen/ — Injection screen: the check that runs on every tool result before an agent reads it — its measured latency against the ceiling it has to stay under.
- /projects/calibrated-decision-layers/saci/ — S.A.C.I.: a typed-decision layer that returns calibrated probabilities and confidence bands instead of prose — method, measurements, and the configurations that failed.
- /projects/calibrated-decision-layers/saci/benchmarks/ — S.A.C.I. benchmarks: fair-comparison rules, the benchmark methodology, and measured rows beside third-party published results on a public decision benchmark.
- /projects/calibrated-decision-layers/saci/log/ — S.A.C.I. log: every shipped change in order, with public corrections kept in place next to the claim they retract instead of being edited away.
- /projects/calibrated-decision-layers/saci/method/ — S.A.C.I. method: typed questions with bounded answers, calibrated probabilities, decision paths, and the fair-comparison rules behind every published number.
- /projects/calibrated-decision-layers/saco/ — S.A.C.O.: which transcript segments an agent keeps at compaction — how filters are scored, why the first label was corrected, and results on future sessions.
- /projects/calibrated-decision-layers/saco/benchmarks/ — S.A.C.O. speed and cost: context-filter latency, model size and training time against the encoders it competes with, and how speed is compared.
- /projects/calibrated-decision-layers/saco/results/ — S.A.C.O. results: the first label measured the summariser. Scores on an independent revisit label, a prospective test the hybrid passed, and a real-path test it failed.
- /projects/crispr-quantum/ — Classical and hybrid quantum-classical networks predicting CRISPR gene dependency from gene copy number (SSCAD 2024), with the full results table.
- /projects/crispr-quantum/method/ — How the SSCAD 2024 paper built and scored its classical and hybrid quantum-classical networks: DepMap data, three architectures, simulated circuits.
- /projects/crispr-quantum/results/ — Full results table of the SSCAD 2024 paper: a classical and two hybrid quantum-classical networks, with every R² negative and the cost of simulation.
- /projects/discount-leverage/ — Discount personalisation as a multi-armed bandit: K-means buyer profiles, a custom gain function, and 3rd place overall at ABI AcademyHack Ambev 2021.
- /projects/local-ai-platform/ — Model serving, bi-temporal memory (750k+ observations) and deterministic agent governance that keep working with the network cable out.
- /projects/local-ai-platform/benchmarks/ — How models are chosen here: a 57-task, 12-domain suite, five verification methods, anti-contamination rules and a pre-registered A/B protocol.
- /projects/roigenius/ — Revenue attribution SaaS in production: eight NestJS microservices, 89 integrated platforms, fixed-point money handling and a dead-letter queue.
- /projects/trading-systems/ — An automated trading program on a demo account: 57 strategies voting in rotation, 8 protection layers, and challengers judged only after 10 days and 40 trades.
- /projects/trading-systems/log/ — Dated capability and rule changes to the demo-account trading program, each taken from a commit, and the phases of its written roadmap.
- /projects/trading-systems/method/ — How the demo-account trading program decides, contains risk and promotes changes: 57 strategies, a 9-phase verdict, 8 protection layers, champion and challengers.
- /research/ — Calibrated decision-layer research: typed questions, calibrated probabilities, ECE-gated benchmarks — including the configurations that failed.
- /uses/ — The software I run daily — mostly self-hosted, and this site carries no third-party scripts, cookies or analytics.
- /api/github-contribs.json — build-time snapshot of public GitHub contributions
- /api/gh-activity.json — build-time snapshot of public activity totals
- /llms.txt — this file

## Conventions

- /now follows the nownownow convention (Derek Sivers): what I would tell a friend I had not seen
  in a year. Not a marketing channel.
- Identity is verifiable via rel="me" links (IndieWeb) and an h-card in the footer.
- Webmentions are accepted at https://webmention.io/marcusnavarrogabrich.com/webmention.
- Theming uses the system preference by default, overridable per-visitor (persisted locally).
- Pages are prefetched declaratively by the browser (speculation rules); no analytics, no beacons.

## Privacy note

Hardware, network topology, self-hosting details, and AI provider specifics are deliberately
excluded from this site. Aggregate telemetry is counts-only (tokens, calls, latency) with no
prompt or provider disclosure. Models are described by class and size band rather than by name, and
the benchmark leaderboard is not reproduced here. Third-party systems named on /projects/calibrated-decision-layers/saci/benchmarks are public
comparison points, not components.

## Contact

- Email: marcusnggg@gmail.com
- GitHub: https://github.com/marcusng8
- LinkedIn: https://www.linkedin.com/in/marcusgabrich

No client-side third-party scripts; no cookies; no analytics service.