Daniel Plas Rivera

2023 — present · Global management consulting firm · Lead Software Architect

AI and LLM platforms

Lead architecture and implementation for AI applications, from model access and retrieval to the APIs and interfaces people use. The work includes an agent tool-discovery layer, a multi-provider model gateway, source-linked responses, offline evaluations, and hierarchical classifiers.

• Built agent tool discovery so clients can retrieve relevant tool definitions without loading the full catalog into every prompt.
• Built a multi-provider model gateway with streaming responses and a shared API for application teams.
• Connected generated answers to source passages, with citation checks and explicit handling of unsupported claims.
• Built hierarchical classification pipelines with reproducible training and evaluation by label and taxonomy level.
• Delivered FastAPI services and React interfaces, including streaming chat and click-to-source citations.
• Built offline evaluation harnesses and mentored engineers through implementation, debugging, and deployment.

Python · TypeScript · FastAPI · React · PyTorch · Hugging Face · Model Context Protocol · RAG · AWS · Terraform · Docker · GitHub Actions

2020 — 2023 · Global management consulting firm · Lead Senior Software Engineer

Data platforms and reliability

Led data engineering for multi-source analytics systems. Replaced repeated API polling with event streaming, built reusable schema checks and pipeline tests, and improved cloud security and operational tooling.

• Replaced repeated API polling with event-driven ingestion using Kinesis.
• Introduced schema validation and automated testing for Glue and Spark pipelines, plus observability and repeatable backfills.
• Hardened cloud infrastructure through encryption, instance metadata controls, security scanning, and Terraform review.

Python · Pydantic · AWS Glue · Spark · Snowflake · Kinesis · Terraform · Docker

2026 · independent research

Fighting-game AI and browser inference

A memorial fighting-game agent trained from a friend’s recorded matches. I work on behavioral cloning and reinforcement learning, recover observations from arcade RAM, and run a trained policy in the browser with a hand-written forward pass.

MLX · behavioral cloning · PPO · VQ-BeT · browser inference

ongoing · independent research

Arcade systems and reverse engineering

I relocate character code inside a 1998 arcade game, write 68000 and Z80 assembly, and build emulator instrumentation to test what the machine actually does. The site includes playable patches, footage, disassembly, and the failed approaches that led to each result.

68000 · Z80 · C++ · Ghidra · Lua · emulation