No day without a line.
Congrats, you just learned a bit about decoding.
About
I was born in Colombia and I moved to London in 2024. Today I work as a forward deployed engineer, helping companies use AI efficiently to solve their problems.
I started in data analytics, then moved into data science and machine learning. With time I did more software engineering, and when the AI boom came I moved fully into applied AI. I’ve worked at big companies and startups, and on consulting projects in the UK, the US and Latin America.
Credits
Roles
- GoogleForward Deployed Engineer2026–now
- Mistral AIForward Deployed Engineer2026
- InformaAI Engineer2024–2026
- Escala 24x7Senior Data Scientist2023–2024
Teaching
- AI MakerspaceAI Engineering Bootcamp, four cohorts2024
Education
- Universidad de los AndesMaster’s, Data Analytics Intelligence2022–2024
- Universidad de los AndesBachelor’s, Economics2016–2020
Projects
All repositories on GitHubTesting TabFM
I compared Google’s foundation model for tables with tuned XGBoost, LightGBM and CatBoost on 51 datasets. It wins on most, but it predicts about 87,000 times slower than XGBoost, so I proposed a fix upstream.
github.com/jjovalle99/ tabfm-eval for Testing TabFM Incident triage agent
Built at a hackathon. It takes a report with text, screenshots or audio, searches the code for the likely cause and opens a ticket in Linear, for about $0.007 per incident.
github.com/jjovalle99/ sre-triage-agent for Incident triage agent Visual RAG with ColPali
A RAG app that reads PDF pages as images, so tables and charts are not lost to text extraction. ColQwen 2.5 finds the right pages and Claude answers from them.
github.com/jjovalle99/ colpali-rag-app for Visual RAG with ColPali Quantizing ColBERT indexes
I applied TurboQuant, a quantizer that needs no training, to ColBERT indexes on three BEIR datasets. At 3 bits they keep 97 to 99% of their quality, but I found it depends a lot on the model.
github.com/jjovalle99/ turboquant-late-interaction for Quantizing ColBERT indexes Text-to-SQL with DSPy
I used DSPy to optimize the prompts of a 7B open model for text-to-SQL. Correct answers went from 51% to 68%, close to GPT-3.5 at 70%.
github.com/jjovalle99/ DSPy-Text2SQL for Text-to-SQL with DSPy notionctl
A Notion CLI for coding agents, on PyPI. It returns compact JSON and clear errors, so an agent can read a spec from Notion and write its results back.
github.com/jjovalle99/ notion-cli for notionctl
Writing
All articles on Medium
Work with me
Mentorship and consulting are my own work, separate from my job. My employer isn’t involved.
Mentorship
For people in tech who want to grow their career.
£90per session
- An in-depth review of your CV
- The tools worth your time
- Staying relevant as AI changes
- Moving countries for your career
Launch price for my first 10 sessions, then £120. Free rescheduling up to 24 hours before. I only give referrals to people I’ve worked with, and I don’t share anything confidential from my job.
Consulting
For companies that want advice or hands-on help using AI in their work.
Fixed pricequoted per project
- Find where AI helps most
- Scope it, then build it
- End-to-end delivery
- Guidance on tools and next steps
It starts with a free 30 min call. Then I send a fixed price and a plan within two working days.
In Latin America? Email me for regional pricing
It’s [14:32] in London.
I reply within two working days.