Jorge Edson

AI & Data Engineer

LLM agents · ML in production · Data pipelines

From raw GPS traces to conversational AI agents — end-to-end systems, shipped.

I'm an AI & Data Engineer in Pernambuco, Brazil. I build systems that run the full distance — raw ingestion to production AI — with a bias for things non-technical users can actually operate.

My flagship is a conversational ReAct agent (LangGraph) that lets non-technical users query a real fleet database in plain language: governed text-to-SQL, RAG, and GPS anomaly detection behind one authenticated API, on cloud or fully-local models. Underneath it is real data engineering — multi-GB DuckDB stores fed by concurrent ingestion — and a collection-coverage / route-optimization stack over OpenStreetMap that's the part of my work I'd call specialist-grade.

• 🤖 AI agents & RAG — governed text-to-SQL, dual cloud/local providers, an MCP server for agent-native tooling
• 🔧 ML in production — a full MLOps spine: train → registry → serve → drift → retrain, live on managed cloud
• 📊 Data engineering at scale — multi-GB DuckDB, 100K+ GPS records/day, hardware-aware concurrent ingestion
• 🗺️ Geospatial & optimization — 4-state coverage modeling and 2-opt TSP routing over OpenStreetMap
• 🎨 Frontend (React / JS / CSS) — a genuine second axis: authored control libraries, hand-built SVG dataviz
• 🌍 Open to remote roles (Brazil & international)

Its own section on purpose — a different licence and a different audience from the showcase below. stillroom is source-available under PolyForm Shield 1.0.0: read it, run it, modify it internally — you may not resell it as a competing product. Source-available, not open source.

A private document assistant that runs entirely on a company's own hardware — ask questions about your contracts, handbooks and price lists in Word, PDF, Excel or CSV, and get answers that cite the document they came from, with nothing sent to any cloud.

• 🔒 No cloud path exists to switch on. There is no fallback provider field, and unknown config keys are rejected rather than ignored — so one cannot be added by a hopeful edit.
• 🙈 The documents are never sent to me either. Indexing runs inside the build on the client's own machine; the only thing that travels back is a readiness report with filenames and question text withheld.
• 🤐 It refuses. When nothing in the documents is relevant, the model is never called — it says so instead of improvising.
• 🔍 The privacy claim is auditable, not promised. The source a client runs is the source published here, so their IT can check what leaves the machine instead of taking my word for it.

Python · FastAPI · ChromaDB · Ollama · Docker · bge-m3 · zero-build UI

Want it set up on your own machines? I take that on as fixed-price work through my Upwork profile: upwork.com/freelancers/jorgeed13. Independent freelancer — not affiliated with, or endorsed by, Upwork.

Four clean-room projects you can browse end-to-end. Most of my production work is employer-confidential, so I rebuilt my patterns from scratch here, with zero proprietary code or data.

The honesty boundary, stated once: every dataset here is synthetic. The numbers below are demonstrations of process — diagnosis, governance, measurement — not product performance. The hosted PdM model is a labelled demo, and nothing here claims a live production deployment. Each repo says the same thing on every surface it appears.

The last two are a pair: a synthetic-telemetry data engine, and the ML-in-production system built on top of it.

A shipped AI agent — ask an invoice ledger questions in plain language, get answers backed by live SQL.

• 🛡️ Governed text-to-SQL — defense-in-depth (read-only connection + allow/deny-list), plus RAG over a collections-policy corpus.
• 🧩 Makes a 0.5B model tool-call reliably — tiny models fail native tool-calling. I measured it, then grammar-constrained the reply to the tool schema: 1/5 → 5/5.
• ⚡ Caches the plan, never the answer — repeat questions reuse validated tool calls, but the SQL re-executes live, so the number is always current.
• ⚙️ AI-native — an MCP server, a Claude Code skill, and a CLAUDE.md: operable by coding agents, not just humans.

LangGraph · FastAPI · React · DuckDB · ChromaDB · Gemini/Ollama · Docker · HF Spaces · MCP

A complete MLOps spine — train → registry → serve → drift → retrain → cloud, running end to end.

• 🚦 A worse model cannot reach production — metric-gated promotion + rollback, asserted by test. Auto-retrain routes through the same gate, so it can never mean auto-degrade.
• 🐛 The bug was in the data, not the model — the classifier scored 0.55 (chance). Rather than tune it, I measured why, traced it upstream, and fixed it there → 0.82 ROC-AUC.
• 🧭 Then I measured that 0.82 is the data's ceiling — three probes converged; one refuted my own hypothesis and I reported it anyway. Knowing the ceiling is the license to stop optimizing.
• 🌩️ Operated on managed cloud, not just containerized — Cloud Run + managed Postgres (Neon) at $0, every prediction logged to the managed DB and read back.
• 🏗️ The infrastructure is code, adopted from a live deploy — the managed stack (API service, a separate worker job, registry, secrets, IAM) is defined in Terraform, imported from the running system until plan came back clean. Writing the IAM down is what proved the permission model was wrong.

MLflow · LightGBM · Optuna · PyTorch · Evidently · Prefect · FastAPI · Docker · Cloud Run · Postgres (Neon)

A synthetic heavy-equipment telemetry generator — where the data is the product. One command builds a reproducible predictive-maintenance dataset for a realistic fleet.

• 📐 Grounded, not faked — signals modeled on the public SAE J1939 standard (real SPNs, units, ranges); climates and road roughness pinned to cited public sources.
• 🧩 The realism most synthetic CAN data skips — capability is gated by model-year era: a signal an old bus never reported is NULL, not zero. Failures degrade progressively toward the event.
• 🔌 Down to the wire — a from-scratch J1939 frame encoder/decoder powers fault injection at the transport layer: corruption and truncation written into the bytes, decoded back as a receiver would see them.
• 🎲 Byte-identical from config + seed. 135 offline tests, ADR-documented.

Python · NumPy · pandas · Parquet · DuckDB · SAE J1939 · pytest

A cross-platform machine-inventory CLI — one double-click scans the whole machine and opens an interactive HTML report.

• 🖥️ 17 self-registering, fault-isolated collectors — CPU, RAM modules, GPU, SMBIOS, disks with SMART, monitors decoded from EDID. A new probe is one module, zero changes elsewhere.
• 🔒 A second, scoped binary whose spec excludes 12 of those collectors — it answers "can this machine run a local AI model?" and cannot read network addresses or serial numbers, which is a property of the artifact rather than a claim about how it was invoked.
• 📦 One binary per OS — PyInstaller, shipped by a tag-triggered Actions release whose per-OS smoke test fails the build if a collector is lost in the freeze.
• 🧪 Standard-library only (psutil the single runtime dependency), 307 offline tests, CI green on Windows/Linux × Python 3.9–3.13.

Python · PyInstaller · GitHub Actions · psutil · pytest · stdlib-only

These are employer-confidential and live in private repos — a walkthrough, not clickable links. The four you can browse end-to-end are above. Happy to do a live code walkthrough of the rest.