Iñaqui Paladino
ML & AI Systems Engineer
About Me
Hi! I'm Iñaqui Paladino Bravo, I'm 24 and I'm from Tandil, Buenos Aires, Argentina.
At 9, I broke my family's router messing with its settings because I needed a static IP to host a Minecraft server. C came next, writing console games. The curiosity hasn't changed, just fewer broken routers.
With an ML system, I want to understand all of it: how the dataset is built, how to tell whether a ranking is good, and what the GPU is doing in the meantime. I'd rather ship a change backed by a measurement than one that just looks better.
I've also built systems where losing data was not an option, which left me with three priorities I bring to everything I build: rigorous evaluation, zero-data-loss architecture and observability.
Experience
AI Engineer
- ❯ Built CONICET's reviewer recommender system, which suggests reviewers for researcher admission, promotion, project and fellowship calls from 27K+ researchers and millions of research outputs integrated from heterogeneous sources, with conflict-of-interest rules modeled in dbt.
- ❯ Designed a decoupled multi-stage Information Retrieval architecture (hybrid Dense + Sparse retrieval on Qdrant, expert-level aggregation, Cross-Encoder reranking and LLM-as-a-Judge), with versioned indexes, zero-downtime deployments and Prefect orchestration.
- ❯ Fine-tuned a Bi-Encoder with LoRA through contrastive learning and knowledge distillation from the Cross-Encoder, on a pair dataset built from the co-authorship graph (Leiden communities) with leakage-free splits.
- ❯ Designed the system's evaluation and auditing: stage-level ranking metrics, feature engineering to audit recommendations, tiered LLM-as-a-Judge and validation against committee decisions for topic classification; experiments, models and presets versioned in MLflow.
- ❯ Developed a conversational agent with LangGraph to explore researchers and publications: intent router with specialized subgraphs, Corrective RAG (query rewriting and relevance grading) and memory with trimming and summarization, keeping data retrieval deterministic via SQL and Qdrant.
- ❯ Created the agent's evaluation suite: synthetic datasets generated from real cases, LLM-as-a-Judge metrics for faithfulness, format and scope, and output guardrails against hallucinated links and internal metadata leakage.
- ❯ Optimized GPU inference with ONNX Runtime (IOBinding), FP16 and multi-GPU batching on ROCm and TensorRT: +140% throughput and −60% VRAM on the Bi-Encoder, 9x on the reranker. Deployed distributed LLM serving with vLLM and HAProxy on multi-stage Docker images.
Software Engineer
- ❯ Automated bank receipt processing, which took ~5 hours of manual work per day, with a serverless event-driven pipeline on AWS (SQS, S3, Lambda, DynamoDB), Textract OCR and deterministic extraction with validation rules.
- ❯ Ensured accounting integrity under concurrency through idempotent processing, conditional writes and DLQ-based retries, with zero data loss.
- ❯ Managed infrastructure as code with Terraform (least-privilege IAM, HMAC-SHA256 signed endpoints) and built observability on CloudWatch: custom metrics, severity-based alarms, and operational and business dashboards.
Machine Learning Intern
- ❯ Built a web platform for Study and Research Paths (SRPs) that groups students' questions by topic and generates reports for teachers; validated in a real classroom pilot and co-authored a paper at CITAD 8 (Barcelona).
- ❯ Designed the unsupervised clustering pipeline (multilingual embeddings, semantic deduplication, UMAP and K-Means with automatic k selection), chosen through EDA and validated against teacher-made groupings.
- ❯ Decoupled analysis from the application (Astro SSR, TypeScript) with an async Python worker and a versioned data contract validated on both ends, all deployed with Docker.
Education & Certifications
Professional Certifications
NVIDIA Certified Associate: Generative AI LLMs (NCA-GENL)
- ❯ Official validation of competencies in LLM architectures, RAG (Retrieval-Augmented Generation), fine-tuning, model evaluation, and deployment of generative models on NVIDIA infrastructure.
AWS Certified Solutions Architect - Associate (SAA-C03)
- ❯ Designing resilient, high-performance, secure, and cost-optimized cloud architectures on AWS. Proven expertise in distributed services, serverless patterns (Lambda, SQS, DynamoDB, S3), and Infrastructure as Code with Terraform.
Academic Background
Systems Engineer
- ❯ Distinction: Graduated with academic honors for excellence in engineering coursework.
- ❯ Overall average GPA: 8.75.
- ❯ Intermediate degree: University Programmer Analyst (APU).
Fundamentals of Software Engineering
- ❯ PILA Virtual Program 2023 (Undergraduate academic exchange).
- ❯ Final passing grade: 9.4.
Projects
Financial conversational agent for market analysis on X. Orchestrated with LangGraph featuring a Corrective RAG (CRAG) loop and hybrid search on Qdrant, integrated with an asynchronous event-driven ingestion pipeline. Deployed on Amazon Bedrock AgentCore with real-time streaming (AG-UI), persistent memory on DynamoDB, Terraform IaC, and end-to-end hardening with Cognito auth, WAF protection, and rate limiting.
Virtual financial analyst built on AWS Serverless architecture. It monitors and analyzes the market on X (formerly Twitter) via a fully decoupled, event-driven workflow. The infrastructure is defined entirely with Terraform (IaC). Orchestration is managed by EventBridge, which triggers ephemeral tasks on ECS Fargate Spot for robust data extraction (handling real sessions and evasion techniques), followed by asynchronous processing on Lambda utilizing Large Language Models (LLMs) for contextual report generation. The frontend operates on a Git-based CMS model: every report triggers a GitHub Actions pipeline that deploys the static site globally via CloudFront.
Serverless solution for ephemeral and secure file transfer on AWS. It includes a CLI to generate temporary upload and download links using S3 Pre-Signed URLs, wrapped in disposable HTML interfaces, with no persistent backend. The infrastructure is fully defined using Terraform (IaC), leveraging remote state and execution via IAM AssumeRole in a multi-account environment, where the Route53 Hosted Zone, the Terraform backend, and ACM certificate validation reside in the management account (AWS Organization). Perimeter security is ensured through CloudFront with Origin Access Control (OAC), keeping the buckets strictly private with TLS/SSL termination.
Fork of the official AWS Labs serverless Dynamic DNS solution for Route53. On top of the original baseline, I added production-oriented improvements for real routers and legacy devices: a DynDNS-compatible endpoint (/nic/update), HTTP/HTTPS support via CloudFront, Basic Auth and temporary token-based authentication for browser updates.
Web-based classroom-style platform designed to support teaching through Study and Research Paths (SRPs), organizing learning around questions and answers. It enables real-time collaboration, academic progress tracking, and automatic analysis of educational data using AI techniques. It is currently used by teachers and students in real classroom contexts, and its development led to a scientific publication presented at an international conference on didactics.
User: usuario@demo.com Password: password
Personal Cloud Infrastructure
Setup and maintenance of a self-hosted cloud infrastructure on a VPS (DigitalOcean), used to deploy and monitor multiple personal apps with Coolify. Includes a private VPN (WireGuard) for secure device communication, a self-hosted S3 server (MinIO) with access control and expiring links, and a N8N instance to orchestrate complex workflows. Also features advanced configuration of firewall, DNS, custom domains, reverse proxy, and SSL certificates.
