Senior AI Full-Stack Engineer

nearShift; Colombie; À distance; COP 380 445 000 - 570 668 000; Plein temps

Avantages offerts par ce poste

Prepaid health plan (Sura)

Birthday day off

Flexible vacation policy

Fully remote with US time-zone overlap

USD 300 referral bonus

Résumé du poste

NearShift is seeking a Senior AI Full-Stack Engineer (remote from Colombia) to help build an AI-assisted SaaS platform for environmental due diligence. You will own the LLM pipeline, grounding, and a React-based report editor that integrates with a Python backend, ensuring verifiable outputs and precise citations.

You’ll work with tools like PyMuPDF and python-docx, design prompts, optimize cost and latency, and lead testing-first development in a distributed, async-friendly team.

Qualifications

• 5+ years of professional software engineering with Python as primary language.
• Hands-on experience shipping LLM-powered features to production with grounding.
• Strong TypeScript + React skills to own editor features end-to-end.
• Experience with document processing: PDF/DOCX, OCR, and extraction.
• Testing-first approach and evidence-driven judgment.

Responsabilités

• Own the LLM pipeline and grounding; design prompts and tool use.
• Build the React editor features and ensure Word round-trip fidelity.
• Write deterministic regression tests and support SaaS readiness.

Connaissances

Python

TypeScript

React

LLM features

Testing-first

Documentation processing

Outils

PyMuPDF

python-docx

pandoc

Docker

Terraform

Hetzner Cloud

Cloudflare

ChromaDB

AWS S3

DuckDB

GitHub Actions

Description du poste

About the Role

We’re looking for a Senior AI Full-Stack Engineer to join a U.S. client building an AI-assisted SaaS platform for environmental due diligence. The platform helps licensed Environmental Professionals produce Phase I Environmental Site Assessment (ESA) reports under the ASTM E1527 standard: it reads the source material behind every assessment — regulatory database reports, historical maps, city directories, aerial photos, and prior reports — and drafts report sections inside a web editor, where the professional reviews, edits, and signs off.

These reports feed real-estate and regulatory decisions, so accuracy is the product. Every claim has to trace back to a primary source, and silently inventing or dropping data is a defect, not a quirk. The platform is in a live alpha with its first professional users and is now being hardened for a wider launch and readied for multiple tenants. This is a senior, hands-on role across the LLM pipeline, the Python backend, and the React editor, working as an embedded member of the client’s engineering team.

What You’ll Own

LLM Pipeline & Grounding

• Build and harden the LLM pipeline. Design prompts, tool use, and orchestration across Anthropic Claude and Google Gemini models, and improve how the platform pulls facts out of long, messy PDFs and spreadsheets and grounds them.
• Make output verifiable. Extend the grounding and citation layer that checks every generated claim against its source document, and decide what happens when a check fails.
• Balance cost and latency against accuracy. Measure token spend and response time, run model bake-offs, and pick the leanest approach that keeps output correct.

Report Editor & Word Round-Trip

• Ship the React editor. Build features in the TypeScript/React report editor — tracked insertions, table editing, accept/reject flows, and evidence review — and connect it to the Python backend.
• Own the Word round-trip. Make sure what the professional sees in the editor matches the .docx delivered to the end client, including formatting, tables, hyperlinks, and highlights.

Quality, Testing & SaaS Readiness

• Write the tests that prove a fix. Deterministic regression tests with stubbed models for every bug, plus integration runs against real reference projects.
• Help with SaaS readiness. Remove assumptions tied to a single customer or data vendor, and support the containerized cloud deployment (Docker, Terraform, Hetzner, Cloudflare).
• Work alongside AI coding agents. The team’s workflow uses Claude Code, Codex, and similar tools: you’ll direct them, review their work rigorously, and own what gets merged.

How You Work

• Accuracy first, then cost. It’s fine to double a model call to remove a hallucination. A 10x cost for a marginal gain is a tradeoff the team flags and discusses before shipping.
• Small, focused PRs. Each one has a ticket, a regression test, and an independent review before merge.
• Written canon. Bugs, priorities, and decisions live in shared documents, so anyone can pick up context without a meeting.
• Real users, real feedback. You’ll see your work used by practicing professionals and iterate on what they find.
• Embedded and async-first. You work inside the client’s team with U.S. time-zone overlap; most collaboration happens in writing — PRs, tickets, and decision logs.

Technical Requirements

Required

• 5+ years of professional software engineering, with strong Python as your primary language.
• Hands‑on experience shipping LLM-powered features to production: prompt design, structured output, tool calling, evaluation, and debugging hallucinations.
• Solid TypeScript + React skills — enough to own editor features end to end.
• Experience with document processing: PDF parsing and extraction (PyMuPDF or similar), DOCX generation and editing (python-docx, pandoc), OCR or image inputs.
• A testing‑first habit: you don’t call something fixed until a test fails on the old code and passes on the new.
• Careful, evidence-driven judgment: you check the source document before trusting the output, and you can say “I couldn’t confirm” rather than guess.
• Clear written English — most collaboration happens asynchronously in PRs, tickets, and written decision logs.

Nice to Have

• Retrieval and grounding systems: embeddings, vector stores (ChromaDB), citation verification.
• DuckDB, AWS S3, or other data-heavy backends.
• Docker, Terraform, and CI/CD with GitHub Actions.
• Building tools for regulated or high-stakes domains (legal, compliance, healthcare, engineering, environmental).
• Familiarity with environmental due diligence, GIS, or real-estate data — not required; you’ll learn the domain on the job.
• Experience leading AI coding agents in a real codebase.

Tech Stack

Python 3.12 · Anthropic Claude & Google Gemini APIs · React 18 + TypeScript + Vite · NiceGUI · DuckDB · PyMuPDF · python-docx / pandoc · ChromaDB · AWS S3 · Docker · Terraform · Hetzner Cloud · Cloudflare · GitHub Actions · pytest

AI-First Engineering Culture

At NearShift, AI isn’t a side project — it’s how we work. In this role AI is both the product and the toolchain: you’ll ship LLM features that professionals rely on, and you’ll direct AI coding agents every day while holding the bar on what gets merged. We look for engineers who use AI to move faster without lowering the standard of proof.

Core Competencies

• Evidence-driven judgment — you verify against the source before trusting the output.
• Ownership — you take a feature from prompt to editor to delivered document.
• Testing-first rigor: a fix isn’t done until a test proves it.
• Clear, candid written communication, including “I couldn’t confirm” when that’s the truth.
• Data-driven tradeoffs between accuracy, latency, and cost.
• Proactivity, continuous improvement, and AI-first thinking.

What We Offer

Compensation & Benefits

• Competitive salary based on your experience and skills.
• Full-time employment contract with NearShift S.A.S. under Colombian labor law.
• Prepaid health plan (Sura).
• A day off on your birthday, plus the Colombian holiday calendar.
• Flexibility in vacations, schedules, and work policies.
• USD 300 referral bonus for every referred candidate who joins NearShift.
• Fully remote work with U.S. time-zone overlap, and access to cutting-edge AI tools and platforms.

Career Growth

• Long-term placement with a U.S. client on a product used by practicing professionals.
• Direct partnership with U.S. engineering and product leaders.
• Real influence on architecture, model choices, and engineering standards.
• A path to grow into tech‑lead and AI‑engineering leadership roles.

Hiring Process

• Intro call (30 min). Get to know each other, the role, and the client.
• Technical conversation (60 min). Walk us through an LLM feature you shipped and how you knew it was correct.
• Paid take-home or pairing session. A realistic document-extraction problem.
• Final conversation with the client team.