Manuel Reyes
Analytics Engineer and Data Engineer, building AI-ready data platforms
I build data pipelines that have to be right, because a wrong number becomes a wrong tax form. My 1099-R reconciliation pipeline runs in production in an ERISA-regulated retirement-plan operation, and I'm rebuilding it in public as a dbt-tested data platform.
Path: business operations → Analytics / Data Engineer (AI-ready data) → Applied AI Engineer → Forward-Deployed Engineer
Proof over keywords
Every project is held to one production standard, and every claim points to code you can open.
Business logic in tested dbt models
A declared grain, a unit test per rule and data contracts at the boundaries. Polars handles ingestion; pandas stays only at the Excel and plotting edges, recorded as an ADR.
Evaluation gates block merges
RAGAS, DeepEval and GEval thresholds are CI conditions, and each metric matches the cost of being wrong. On PostCheck the gate is the false-alarm rate, not accuracy.
Reproducible and enforced
Python 3.14 pinned once, a committed uv.lock, and pre-commit hooks that are a strict subset of CI. Secret scanning and notebook stripping make synthetic-data-only mechanical.
Observable and PII-safe
One structlog chain renders my logs and my dependencies', with a redaction processor inside it, so masking never depends on remembering it.
In production: the 1099-R reconciliation pipeline
Reconciles retirement-plan distributions across two recordkeeping systems, derives the expected 1099-R Box 7 code, and produces correction files for human review.
Now: the rules are being rebuilt as dbt models with a unit test per rule, and the shipped pandas version stays frozen as the baseline the rebuild is measured against. Read the code
Public repos use synthetic data only. Claims are limited to mechanism and relative change: no participant, plan, client or employer-identifying data.
Portfolio
Four flagships and two supporting projects, ordered by the roles I'm targeting. AI is powered where the model does the core job and assisted where a deterministic core must make the decision.
-
Flagship for analytics and data engineering
The live reconciliation pipeline, hardened into a governed data platform: dbt-tested models with a declared grain, data contracts, Airflow, a Docker/ECS deploy, monitoring and a semantic layer. An LLM explains exceptions and, in Stage 3, proposes corrections; a human approves every write.
- Production live in a regulated operation; success rate held against a freshness SLA
- Cost 95% less weekly effort; relative figures only
- Architecture ERISA-driven ADRs, C4 context and container views, dbt tests and contracts
Python, SQL, dbt, Polars, Airflow, Snowflake, Great Expectations, Docker, AWS ECS, Terraform
-
Flagship for data engineering and applied AI
Reviews posted retirement-plan distributions against the request packet and a versioned SOP. Each packet is claimed exactly once, the posted export is parsed deterministically, and a multimodal model reads the scanned forms; rules run first and the LLM handles only what's left. Every verdict lands in an append-only event log that feeds a dbt quality mart.
- Production read-only by a tested invariant; a verdict is a recommendation, never an authorization
- Cost one eval suite run on Anthropic, Azure OpenAI and local Ollama
- Architecture declared grain (one row per packet per review run), ADRs, C4
Measured against a prompt-only reviewer already in use. The blocking gate is the false-NIGO rate, not accuracy.
Python, Anthropic SDK, Azure OpenAI, Ollama, Pydantic, dbt, DeepEval, FastMCP, Docker
-
Flagship for applied AI
Answers policy questions with cited sources and escalates when it isn't sure. It starts as RAG over ChromaDB and grows into GraphRAG with Neo4j, with access control enforced at retrieval time.
- Production RAGAS and DeepEval gates are merge conditions; exposes a FastMCP server
- Cost cost per query and p95 latency across local and cloud models
- Architecture retrieval ADRs with rejected alternatives; C4 context and container
Python, Anthropic SDK, ChromaDB, Neo4j, FastMCP, RAGAS, DeepEval, Streamlit, Docker
-
Flagship for applied AI and execution safety
Takes a trading strategy from backtest to paper to live. An LLM analyst reads aggregated backtest results and proposes hypotheses; a deterministic core owns every order, and the live path requires human sign-off and has a kill switch.
- Production sealed out-of-sample vault, walk-forward validation, intended-vs-filled reconciliation
- Cost compute per backtest sweep and data-feed cost
- Architecture execution and risk ADRs, C4, a shared point-in-time primitives library
Research project. Not investment advice, and no claim of positive returns.
Python, NautilusTrader, DuckDB, Parquet, Optuna, Ollama, LangGraph, Alpaca, DeepEval
-
Supporting project, small-cap research
An LLM reads SEC filings under a faithfulness gate (≥ 0.9), and an ML model on a point-in-time event study predicts which trigger events follow through, reported as calibrated lift with false-discovery control, never as returns.
Python, edgartools, DuckDB, scikit-learn, LightGBM, Anthropic SDK, DeepEval
-
Supporting project, document AI
Reads scanned distribution forms with a multimodal model, validates every field against a frozen Pydantic contract, and routes low-confidence forms to a person.
Python, vision LLM, Pydantic, DeepEval (GEval), Streamlit, Docker
Path: stage 1 of 3
Stages end on evidence, not dates.
-
Stage 1: Internal AI Builder
Active, months 1–8
Production Python, SQL and dbt, and evaluation-first AI. The name describes the build work, not a job title.
Ends when DataVault Stage 2 ships. Exam: Azure AI Fundamentals (AI-901).
-
Stage 2: Analytics / Data Engineer, AI-ready data
Target role, months 9–20
dbt on Snowflake, Airflow, data contracts and a Docker/ECS deploy, plus AI-ready data: embedding pipelines, PII-safe corpus preparation and point-in-time datasets.
Exams: Fabric Data Engineer (DP-700), AWS Data Engineer Associate.
-
Stage 3: Applied AI Engineer to Forward-Deployed Engineer
Months 21–32
Production agentic systems, MCP, human-in-the-loop gates and evaluation as engineering, plus customer discovery and problem decomposition.
Exams: Azure AI Apps & Agents (AI-103), NVIDIA NCA-GENL, Databricks GenAI Engineer, Neo4j.
Degree track: B.S. in Industrial & Systems Engineering; CU Boulder graduate algorithms pathway from November 2026; Georgia Tech OMSCS (Computing Systems) next.