| status | canonical |
|---|---|
| authority_scope | builder-story-and-case-studies |
| owner | Aaron Marchant |
| last_verified | 2026-09-02 |
| verified_against_commit | 4e729c2 |
| claims_source | docs/CLAIM_LEDGER.md |
| supersedes | |
| superseded_by | |
| archived_at |
AI Systems Engineer & Solo Founder
Creator of REI.ai — The OpenAI-Compatible FinOps Proxy & Pre-Spend LLM Router.
"REI does not begin by asking for control of your AI traffic. It begins by earning the right to recommend a change."
- Inference FinOps & Pre-Spend Routing: Deterministic pre-flight model selection, OpenAI-compatible proxy gateways (
/v1/chat/completions), prompt-freeze caching, and 3-bucket traffic audits (97.3502% measured input-cache hit rate across 1,848,473,560 tokens). - Evidence-Bounded Decision Audit: Replay analysis precedes live changes. The isolated
ExecutionControllerunit contract preserves the requested model and adds no provider call in shadow mode; production integration remains a separate gate. - Adversarial Security & Local Model Gates: A 16-category D1 taxonomy, fixed red-team regression corpus, Feynman Gate evaluation harness, and an incomplete local-model evaluation retained with its failure evidence.
- Full-Stack AI Engineering: End-to-end React/TypeScript interfaces, serverless backends, Customer Pilot Workspace (
/#pilot), and hexagonal multi-package runtimes (EchoForge). - Empirical Rigor: 1,366/1,366 automated tests across 121/121 suites passed locally on 2026-09-02, with 1,028 commits on
main.
┌─────────────────────────────────────────────────────────────────────────────┐
│ THE 3-PILLAR TRIAD │
├─────────────────────────────────────────────────────────────────────────────┤
│ 1. REI.ai ──► AI Systems & FinOps (Proxy, Evidence, 1,366 Tests) │
│ 2. Arena Harness ──► AI Security & Evals (Feynman Gate, 136 Blind Prompts)│
│ 3. Family Archive ──► Full-Stack Product (GPS Evidence Tiers, Provenance) │
└─────────────────────────────────────────────────────────────────────────────┘
Repository: github.com/aaronmarchant96-max/rei-ai · Live: https://rei.ai
- Problem: Teams lack evidence about which routine requests may be candidates for less expensive models without degrading task-specific quality.
- Architecture:
- OpenAI-Compatible Gateway (
/v1/chat/completions): Drop-in proxy for Cursor, Cline, Aider, and backend pipelines. - Pre-Spend Selection: Uses deterministic in-memory policy without calling an LLM to route an LLM. A fresh retained benchmark is required before publishing a numeric latency ceiling.
- 3-Bucket Audit Segmentation: Categorizes requests into Candidate to Shadow, Retain Current Tier, and Insufficient Evidence.
ingestable ≠ replay-routable: Missing or redacted prompt text is normalized in denominator audits but excluded from savings claims.- BYOK SaaS Model: Customer-owned provider keys; zero inference balance-sheet liability.
- OpenAI-Compatible Gateway (
- Measured Telemetry:
- 1,366/1,366 passing automated tests across 121/121 suites in the latest local run.
- 1,028 total commits on the
mainbranch as of 2026-09-02. - 1.848B development tokens processed through OpenCode/DeepSeek build workflow for $23.52 ($567.06 savings vs $590.57 no-cache counterfactual).
- 70.6% pooled classification accuracy (96/136 unique samples), with implemented-route holdouts ranging from 90% to 100% under their documented exclusions.
- Reproduce from Clean Checkout:
git clone https://github.com/aaronmarchant96-max/rei-ai.git cd rei-ai && npm install npm test npm run dev
Integrated Module: rei-ai/src/__eval__ · docs/DEFENSE_IN_DEPTH_CONTROL_MATRIX.md
- Problem: AI benchmarks often suffer from dataset contamination, brittle regex parsers, and ungrounded claims. Teams lack standardized ways to test model resilience against prompt injections, system extraction, and quality degradation.
- Architecture:
- D1 Threat Taxonomy: 16-category zero-token scanner flagging recursive jailbreaks, base64 ciphers, credential leaks, and identity spoofing before API dispatch.
- Feynman Gate Suite: 136 ground-truth holdout queries evaluating accuracy across 5 specialized reasoning domains.
- Local Model Quality Gate: Evaluates local candidate models (LLaMA 3.2 3B) separating CARDO structural score from Epistemic correctness score.
- Measured Result:
- 12/12 correct routes on the fixed red-team regression corpus, which exercises 11 taxonomy categories; a separate five-entry replay measured 75% route adherence.
- The local-model overnight run is incomplete (98/136 records, including one delivery failure) and is not represented as promotion evidence.
- Tool arguments are covered by JSON/Zod schema-validation and retry tests; no universal zero-failure claim is made.
Repository: github.com/aaronmarchant96-max/family-archive · Engine Spec: docs/FAMILY_ARCHIVE_PORTING_SPEC.md
- Problem: Historical databases suffer from catastrophic hallucination when AI systems merge records of individuals sharing identical names and birth years.
- Architecture:
- 4-Tier Genealogical Proof Standard (GPS) Classifier: Enforces strict epistemic tiers (
primary_direct,secondary_derivative,inferred_modeled,negative_search). - Disambiguation Hinge Evaluator: Isolates conflicting facts before asserting identity matches.
- Negative Search Audit Receipts: Logs exhaustively searched databases where no record was found.
- 4-Tier Genealogical Proof Standard (GPS) Classifier: Enforces strict epistemic tiers (
- Measured Result:
- Citation and provenance requirements are enforced through schemas and integrity tests; this is not a claim that every generated assertion has been externally audited.
- Reusable standalone TypeScript library (
archivistEngine.ts) with dedicated unit test suite.
┌─────────────────────────────────────────────────────────────────────────────┐
│ AARON'S BUILD BENCHMARKS │
├─────────────────────────────────────────────────────────────────────────────┤
│ • Total Tokens Processed ──► 1.848 Billion development & evaluation tokens │
│ • Total Build Spend ──► $23.52 API spend (97.35% input cache hit rate) │
│ • Verified Test Suite ──► 1,366/1,366 tests across 121/121 suites (local) │
│ • Git History ──► 1,028 commits on main (measured 2026-09-02) │
│ • Operating Budget ──► ~$60/month lean serverless infrastructure │
└─────────────────────────────────────────────────────────────────────────────┘
- 🌐 Live Platform: https://rei.ai
- 📦 Repository: github.com/aaronmarchant96-max/rei-ai
- 🐦 Twitter / X: @PromptHound96
- 💻 GitHub: github.com/aaronmarchant96-max

