hello, Jair herehello, Jair here|

Cipli Research Center Sociedad Anónima Cerrada | FractalUp ↗

Full-time · Lima, Peru · Remote

1. GenAI2026.05 — Present+

   • Python
   • Pydantic AI
   • FastAPI
   • SurrealDB
   • FalkorDB

   01RAG pipelines & indexingChunking, embeddings and incremental vector index sync.+

   Ingestion of heterogeneous sources (text, conversations, transcripts), chunking, embedding generation and incremental index sync to keep context fresh without full re-indexing.

   • Python
   • FastAPI

   02Hybrid retrievalDense + BM25 search and semantic search to retrieve relevant context.+

   Retrieval strategies combining dense vector search (HNSW), BM25 and semantic search with metadata filters, to improve the precision of the context given to the model.

   • SurrealDB

   03Agents & structured outputsPydantic AI agents: tool calling, validated outputs and Logfire tracing.+

   Pydantic AI agents orchestrating tools over multiple knowledge sources, with schema-validated typed outputs, served with FastAPI and observed with Logfire / OpenTelemetry.

   • Pydantic AI
   • FastAPI
   • Python

   04Multimodal summarizationWindow-based summarization (time and volume), combining OCR and vision models.+

   Summarization pipelines triggered by time windows and data volume, with multimodal input: OCR text extraction and direct image analysis with vision models.

   • Python

   05Knowledge graph + vectorsMigrated from FalkorDB to SurrealDB (graph and vectors in one engine) and trained the team.+

   Knowledge modeled as a graph (nodes and edges) and migrated from FalkorDB to SurrealDB to unify graph and vectors in one engine, simplifying retrieval. Trained new team members on graph modeling, edges and vector search.

   • SurrealDB
   • FalkorDB
2. Desarrollador back-end2025.11 — Present+

   • Elixir
   • Phoenix
   • Valkey
   • OVH Object Storage

   01Idempotent event ingestionHigh-frequency events with deduplication and a cost-efficient design.+

   High-frequency event processing (every few seconds per client) with idempotency keys to tolerate retries and duplicates, plus presence-state computation. Designed without per-entity processes (GenServers) to scale at low cost.

   • Elixir
   • Phoenix

   02Real time & cachingMigrated from GraphQL to WebSockets (Phoenix Channels) with Valkey caching.+

   Migrated a GraphQL API to WebSockets with Phoenix Channels, scheduled jobs and push notifications, using Valkey as a cache to reduce database load.

   • Elixir
   • Phoenix
   • Valkey

   03Object storage migrationFrom Backblaze B2 to S3-compatible storage (OVH), including legacy code.+

   Migrated data and integrations from Backblaze B2 to OVH Object Storage through its S3-compatible API, updating legacy implementations.

   • OVH Object Storage
   • Backblaze B2