AI Engineer · M.Sc. in Systems and Computing Engineering (AI) · Pereira, Colombia
I read AI papers, rebuild them from scratch, and then test them the way they'd actually be used, which is rarely how the benchmark was built. Then I ship them as working systems on a single consumer GPU.
My master's thesis was about when not to trust a model: whether the explanations behind a graph neural network's money-laundering alerts are stable, plausible and faithful. Mostly they weren't, and the three didn't move together. I've kept asking that question since, in every domain I touch.
- A benchmark result can halve under realistic testing. I reproduced LogBERT's log anomaly detection (F1 83.2 vs. 82.3 published). Trained on the past and tested on the future, the way it would be deployed, it dropped to 38.
- A one-line rule beat a transformer on the same benchmark: "flag any log message never seen in training" scored 96.8 against LogBERT's 88.1.
- Reading the customer's history is worth 30 points of F1, and a year of drift erases all of it. On 24M card transactions, a transformer over the last ten purchases reached F1 0.89 against 0.59 for a per-transaction XGBoost. Trained on the past and tested on the next years, every model fell below 0.08: fraud comes in new campaigns.
- An LLM agent couldn't plan a port, and was the best part of the system anyway. With Reflexion, a 4B model improved on HumanEval (83.5% to 88.4%), yet produced 0 executable berth plans in 30 scenarios that an exact solver optimized in under half a second. As the operator's interface it understood 93% of requests in English and Spanish, and letting code do the arithmetic removed its addition mistakes.
- "Written for a 5-year-old" is a claim you can measure. Asked for a children's story at a given age, a 4B model landed at the right Spanish reading level (Fernández-Huerta plus sentence length) 42% of the time; measuring each story and having it rewrite the sentences that missed brought that to 71%. Along the way I reproduced the three claims of Adversarial Diffusion Distillation (SDXL-Turbo) on one RTX 4060.
- A detector is only as good as the country it was trained in. RT-DETR beat YOLO on road damage in every country I tested (0.57 vs 0.50 mAP50 in Japan) and ran at 95 FPS after TensorRT; trained on Japan alone, YOLO fell from 0.49 to 0.07 in India, and on a country no model had seen, all of them dropped to 0.08–0.16.
- A time-series foundation model lost to the calendar. Chronos reproduced its paper to the third decimal, but on daily Colombian rain it scored worse than "what usually falls this week of the year". Feeding its forecasts into a landslide risk model still ranked days better than that baseline, because most of the signal was the rain that had already fallen.
- Small specialists beat big generalists where it counts. A 150M-parameter biology model identified Colombian birds at 80.5%; an 8B vision-language model managed 30.5%, while claiming 95% confidence on 184 of 200 photos.
- LLMs invent unless the system forbids it. Asked why a Bitcoin transaction was flagged, a local LLM made up specifics in 30 of 30 answers. Forcing every claim to cite a fact, and checking it with code, got 90% of claims grounded.
- Retrieval changes what an agent says, not just how sure it sounds. From memory, a security agent labelled SSH brute force as "Valid Accounts" every time; with the MITRE ATT&CK catalogue retrieved, it named Brute Force in 63% of cases.
- A popular explanation method draws the image, not the model. On chest X-rays and blood smears, Guided Backprop's heat map stayed 91–93% the same after I randomized the network's top layers. Gradient maps passed that test but changed 95% of what they highlighted under invisible noise.
- A stable explanation isn't a faithful one. 56% of GNNExplainer's explanations changed when re-run with another seed, and the edges they highlighted often didn't move the model's score at all.
- Evaluation that reflects deployment: temporal splits, base rates, calibration, abstention, cost per prediction
- LLM systems that cite, check and defer: GraphRAG, agents with human approval, LLM-as-judge done carefully
- Explainability that's measured rather than drawn: stability, fidelity, plausibility
- Making models cheap enough to actually run: small specialists, ONNX, quantization, local inference
Before AI I spent years building software end to end: Python and Java backends, React and Angular frontends, and data platforms on AWS and Azure. That's why I care about the parts around the model: APIs, CI/CD with quality gates, monitoring, and what happens when the data drifts.
Toolbox: PyTorch · PyTorch Geometric · Hugging Face · OpenCLIP · LangGraph · Ollama · FastAPI · ONNX Runtime · diffusers · Docker · Jenkins · Terraform · MLflow · DVC · Langfuse · React · Next.js · Angular · AWS
Why your model's validation score won't survive production · graph ML for fraud · getting an LLM to say "I don't know" · Colombian biodiversity and data