Manuel Parra
Senior AI Engineer | OCR, RAG & FastAPI | Applied AI, backend, and platform systems
Contrata sus servicios
Solicita un presupuesto sin compromiso
Sobre mí
Lead software engineer focused on applied AI, backend, and platform systems. Builds source-grounded LLM workflows, document-processing pipelines, evaluation paths, and model-serving systems, with production experience in fintech, identity, consent, and distributed services. Works hands-on across architecture, implementation, testing, cloud delivery, and incident investigation.
Document AI and OCR: document data extraction and validation with Python and FastAPI, including source references and human review.
AI agents and RAG: semantic search over approved sources, with citations and evaluation for assistants that answer questions about documents.
Backend and AI workflow automation: API development with FastAPI and API integration for document workflows and business systems.
Model serving and LLM deployment: cloud delivery and vLLM inference investigation.
Remote / UTC-4
Spanish (native) / English (C1)
Experiencia
Professional experience
Technical Lead (IC) | ConTrust Suite - Construction Daily Reporting PlatformMarch 2026 - Present
- Designed and coordinated a multi-repository, tenant-aware reporting platform spanning FastAPI services, React/Vite workflows, PDF workers, shared schemas, analytics, and AI-assisted reporting. Ingestion flows preserve raw values, validation flags, source references, and document evidence.
- Improved reliability with source-evidence contracts, idempotent workers, real-document regression tests, accuracy benchmarks, and retry and concurrency controls. Delivered seven review, reporting, analytics, and grounded-question workflows.
January 2026 - March 2026
- Led a three-engineer build of a contractor prequalification, KYB, and financial-risk platform. Built the FastAPI/PostgreSQL document path for Arabic and English OCR, vision extraction, deterministic accounting checks, source evidence, and human review.
- Improved an audited-statement extraction evaluation from 34% to 94%. One 22-page filing processed in 97 seconds at roughly USD 0.30 to USD 0.42.
October 2025 - December 2025
- Built Rust services for identity-based encryption, BLS signatures, Merkle proofs, threshold decryption, and consent verification. Secured gateway flows with provider isolation, OIDC/JWKS, API-key rotation, revocation-aware consent, and restart-safe recovery.
- Deployed asynchronous threshold-node infrastructure with NATS JetStream, encrypted key-share storage, deterministic coordination, health checks, metrics, Railway, and Podman.
March 2024 - October 2025
- Owned backend and platform architecture for a fintech product using Node.js and TypeScript. Integrated Plaid OAuth, Bridge.xyz KYC/KYB, virtual accounts, transfers, webhooks, idempotency, reconciliation, and scheduled workers.
- Processed more than USD 500K through linked virtual accounts over six months with no integration incidents. Automated Aave V3 risk controls across 84 positions and more than USD 2M in collateral flow, with zero managed-position liquidations over 18 months.
September 2020 - March 2024
- Built REST APIs, MongoDB schemas, WebSocket and Socket.IO state flows, external sports-data integrations, validation, OpenAPI documentation, and deployment setup.
Selected AI and platform work
Evidence-grounded portfolio agent | LangGraph / Next.js | 2026- Built an agent with approved-evidence retrieval, citation checks, bounded context, cancellation, and explicit completion states. On nine fixed questions, hybrid retrieval raised Recall@6 from 88.9% to 100% and MRR from 0.704 to 0.833.
- Implemented and locally verified a strict stateless OpenRouter MCP server for the 2026-07-28 protocol, with no initialization or session fallback. A shared operations layer powers 13 MCP tools and a direct CLI, with separate management credentials, bounded analytics, 21 deterministic tests, and live public API contract checks.
- Built a security-hardened Windows derivative of Farza's Clicky with push-to-talk, screen-aware answers, visible-element pointing, multi-monitor and mixed-DPI routing, speech modes, and explicit privacy controls.
- Built a multi-tenant job-search platform with six AI agents for role matching, application coaching, sourcing, tracking, and job discovery. Used FastAPI, React, PostgreSQL row-level security, encrypted document storage, and OIDC-based authentication.
- Diagnosed repetitive output from a 504-billion-parameter checkpoint by tracing the sparse layer-indexing path in vLLM. Restored correct output and validated cold prompts up to 240K tokens.
Core skills
Applied AI: LangGraph, LLM agents, MCP, RAG, hybrid retrieval, structured outputs, evaluation, OCR, multimodal models, vLLMBackend and Data: Python, FastAPI, Pydantic, SQLAlchemy, Alembic, TypeScript, Node.js, NestJS, REST APIs, PostgreSQL, MongoDB, Redis
Distributed and Platform: Event-driven services, idempotency, retries, RabbitMQ, NATS JetStream, SSE, WebSockets, Docker, CI/CD, GitHub Actions, GCP Cloud Run, Cloud SQL, Secret Manager, Artifact Registry, Cloud Build, IAM, IAP, AWS, Azure
Security and FinTech: OIDC, JWKS, OAuth, key rotation, provider isolation, audit logs, access control, KYC/KYB, Plaid, Bridge.xyz
Educación
Bachelor's in Computer Science
Universidad de Carabobo
2015 - 2021




