Senior AI Engineer - Tech Lead

Santex is a global Certified B technology company as well as one of North America's fastest-growing companies, according to the Inc. 5000. It’s present in 18 countries and over 100 cities, and headquartered in Argentina with offices in the United States, Mexico, and Peru.

Its two service lines — AI Tech Consulting and AI-Optimized Engineering — help organizations comprehensively adopt technology, expand their capabilities, and improve business outcomes at scale.

The same approach defines its culture: Santex has 80.4% employee engagement (global average: 21%), 8% annual turnover (tech sector: 20–25%), 9,369 training hours invested in 2025 — more than double the global average —, and a 26.5% year-over-year reduction in carbon footprint, achieving carbon neutrality through offsetting.

Santex. A company with purpose and judgment. Yes, it's possible.

We are an Equal Opportunity Employer and are committed to fostering an inclusive and diverse workplace. We do not discriminate on the basis of race, color, religion, sex, gender identity or expression, sexual orientation, national origin, age, disability, veteran status, genetic information, or any other characteristic protected by applicable law. All qualified applicants will receive consideration for employment without regard to any of these factors. We strongly encourage candidates from all backgrounds to apply.

Job Description: We are looking for a Senior AI Developer to join our team. The ideal candidate will have extensive experience in developing and deploying AI and machine learning models.

About the role
Over the next 14 months we're replacing Solr with an AI-native retrieval and generation platform on AWS Bedrock + OpenSearch, and rolling out 9 AI use cases on top of it — AI Augmented Search, Chair Assistant, Ask member-facing, Meeting Prep, Topic Tagging, Content Customization, Digest Emails, People-Like-Me, and a Strategic Planning Wizard.
You're the technical lead for everything that touches retrieval and generation. You set the patterns in Foundation, then carry them through the eight feature phases. The choices you make in the first 16 weeks — chunking strategy, hybrid retrieval design, evaluation harness, prompt registry — propagate to every downstream use case. We need someone who has been here before.
You'll work alongside an external architecture partner (Santex Lab) during Phase 1 for knowledge transfer. By Week 17 you own the platform.

What you'll own
Retrieval architecture
Hybrid retrieval design on OpenSearch (BM25 + k-NN) with RBAC pre-filters across 28 content types.
Per-content-type chunking strategies (paragraph, section, thread-aware, comments-as-array). Tier-A content types include articles, best practices, PeepSo discussions, calendar events, and LMS lessons.
Embedding pipeline on Bedrock (Titan v2 default). Batch + incremental flows. Multilingual fallback path.
Index design — content / profile / metadata indices, mapping per index, k-NN tuning.

Generation layer
Prompt registry — versioned per use case, rollback-ready, with per-field prompts for the SP Wizard (~60 fields).
LLM routing through Bedrock — model selection per intent tier (name lookup, navigation, factual, RAG, advisory).
Bedrock Agents for Chair Assistant + member-facing knowledge bases. Multi-turn orchestration, moderator handoff.
Bedrock Guardrails configuration — PII filters, denied topics, sensitivity classifications.

Evaluation and observability
Search-quality validation dataset (200–500 query/answer pairs seed → ongoing growth). Regression harness run on every prompt / model / index change.
RAG eval metrics — retrieval@k, answer faithfulness, citation accuracy, hallucination rate, suppression precision.
A/B testing framework — cohort assignment, metric collection, toggle infrastructure for safe rollout.
Prompt/response capture for Product team visibility (a non-negotiable Phase 1 deliverable).

Cross-cutting
Cost guardrails — monthly ceilings, model downgrade triggers, throttle rules. The stated cost ceiling is ~$10–18k/yr at 10x current Solr volume; the design must hold.
Knowledge transfer — co-author ADRs, document patterns across five dimensions (what / why / operate / extend / debug), pair-program with mid engineers so the team owns the reasoning, not just the result.
 
Must-have qualifications
5+ years backend / ML engineering experience. 2+ years of which is shipping production RAG systems serving real users — not POCs, not demos.
Hands-on experience with at least one production vector store (OpenSearch k-NN, Pinecone, Weaviate, pgvector, Vespa, or comparable). OpenSearch experience strongly preferred.
Production embedding pipelines — chunking strategy design, batch + incremental ingest, dealing with the realities of mixed content types.
Hybrid retrieval (BM25 + vector) with proper score fusion. You've tuned k-NN weights against eval datasets and know why arbitrary thresholds don't survive contact with real corpora.
Prompt engineering for production — versioned prompts, regression testing, model swaps without breaking downstream consumers. You've designed a prompt registry before, or you can clearly explain how you'd build one.
RAG evaluation experience — you've built or run an eval harness with retrieval-quality + generation-quality metrics. You understand why "vibe check" doesn't scale past a handful of queries.
AWS fluency — IAM, Lambda, SQS, API Gateway. You don't need to be a platform engineer (we have one), but you can read CloudFormation/CDK and ship code into the AWS environment without holding hands.
Strong written communication. You'll write specs, ADRs, runbooks, and review code from mid engineers. The team transitions to internal ownership — the docs are the deliverable as much as the code.

Strongly preferred
AWS Bedrock production experience — model invocation, Knowledge Bases, Agents, Guardrails. If you've shipped on Bedrock specifically, it cuts months off ramp time.
OpenSearch Managed ops experience — index design, cluster sizing, hybrid query construction.
Compliance-aware engineering — SOC2, US-only data residency, RBAC enforcement at the retrieval layer. Member data is "proprietary and confidential" — public model training is a hard no, sensitive-content boundaries are real.
PII handling — guardrail configuration, sensitive field exclusion, draft/unpublished content suppression.
Experience replacing a legacy search system (Solr → modern stack). Parallel-run cutover patterns, traffic shifting, deprecation triggers.
Multi-content-type retrieval — community posts + LMS content + structured forms + attachments. You've designed an index that serves heterogeneous content without collapsing into a least-common-denominator schema.

Nice to have
Bedrock Agent product experience including current limitations and when to fall back to custom orchestration.
PDF/image extraction at scale (Textract or comparable) — we have ~6.4K attachments in the corpus.
Multilingual retrieval — TranslatePress is active across 13 languages.
Experience with PEEPSO, LearnDash, or WordPress-based content platforms (we don't expect this — we have a WordPress engineer).
Background in a regulated or enterprise-content domain (legal, healthcare, financial services).
 
Advanced English:
Excellent verbal and written communication skills in English.