● Irving Ernesto Quezada Ramírez / CTO @ RoomIQ Xalapa, Veracruz, MX
AI engineer,
physicist.
I build machine-learning systems and the infrastructure that keeps them alive: from fine-tuned LLMs for endangered languages to data platforms moving 80,000 listings a day.
01 / Selected work
Things I've built.
- 01 RoomIQ CTO & Co-founder Market intelligence for room-rental investors. 600+ US cities, refreshed daily. 2023 → Now
- 02 SteelEye Main Developer An operating system for steel contractors, with ML that reads construction blueprints. 2025 → Now
- 03 NeuralTranslate Creator & Researcher An open Nahuatl to Spanish machine-translation model, and the lesson that the data mattered more than the model. 2025 → Now
- 04 Flight Data Infra Data Engineer · FindMyFlight.ai One API, six airlines, six different locked doors. The data platform behind FindMyFlight.ai. 2025
02 / About
I'm a physicist who fell in love with systems that learn, and the unglamorous engineering that makes them actually work.
By day I'm the CTO of RoomIQ, where our pipeline ingests 80,000+ rental listings daily across 600+ US cities and turns them into revenue forecasts investors can act on. I'm also the main developer of SteelEye, an operating system for steel contractors where I trained ML models that read construction blueprints, detecting and naming structural members straight from CAD vector data.
Before that: a physics degree (3.9 GPA), GPU-accelerated research pipelines, and a habit of reverse-engineering things that weren't meant to be reverse-engineered: enterprise bot protection, mobile APKs, and airline APIs.
On the open side, I publish models and datasets for indigenous and low-resource languages (Nahuatl, Quechua, Akkadian) as Thermostatic on HuggingFace.
Experience
- 2023 → Now CTO & Co-founder · RoomIQ Market intelligence for room-rental investors across 600+ US cities.
- 2025 → Now Main Developer · SteelEye The operating system for steel contractors: estimating, job costing, AIA billing.
- 2025 → Now Data Engineer · FindMyFlight.ai Airline-data infrastructure. Recruited after winning a scraper performance contest.
- 2025 → Now Software Engineer · Portals (UK) Real-time LLM localization for enterprise live streaming in 20+ languages.
- 2024 → 2025 Research Assistant · Universidad Veracruzana · CEICAH GPU-accelerated time-series analysis over YOLO detection pipelines.
n-body · barnes-hut · θ = 0.7 · n = 4000
03 / Open source
In the open.
codecontexter
Lightning-fast CLI that packs a whole codebase into one LLM-ready file. Respects .gitignore, redacts secrets.
camoufox-mcp
MCP server that lets AI agents drive an anti-detect Firefox: fingerprint spoofing, request capture, human-like input.
tlshttp
Modern wrapper for tls-client with an httpx-like API, browser TLS fingerprinting, and async support via anyio.
jspace-refusal
Interpretability research: locating and editing refusal directions in neural networks via Jacobian analysis.
neblinia-speech
Foundational ASR for Mexican indigenous languages, trained with GSPO reinforcement learning.
ppocrv6-fast-cpu
Lossless CPU inference for PP-OCRv6 detection + recognition through OpenVINO. OCR without the GPU bill.
04 / Writing
Field notes.
The 0.98 F1 That Wasn't: Catching Your Own Benchmark Cheating
An eval-integrity story: how a 0.987 member-F1 turned out to be a model grading its own homework, and the general rule for measuring the thing you think you are measuring.
GSPO: Reinforcement Learning for Low-Resource Speech Recognition
Teaching Whisper to hear Mexican Indigenous languages with RL. Hand-rolling GSPO around an audio policy, rewarding on negative Character Error Rate, and making real progress on languages every off-the-shelf system fails.
Reading Refusal Before the Model Speaks
An interpretability study with the Jacobian lens: the model commits to refusing roughly ten layers before it writes a word, that decision is legible in the verbalizable workspace, and a surgical pullback edit removes only about a third of it. Where the rest of the 'no' actually lives.