Projects

Chasing RSVPs is one of the most tedious parts of planning an event: hundreds of messages, guests who answer in their own words, and a spreadsheet that’s always out of date. Konfirmo hands that work to an AI. Planners launch a WhatsApp campaign to their guest list, the assistant understands each reply, and confirmations land in one place automatically.

The problem was information asymmetry. People didn’t know what their insurance covered, and agents couldn’t answer them fast enough. Questions like “does my policy cover this?” or “which hospitals can I go to?” waited in a queue for a human.

The assistant lives in WhatsApp. A user can ask about coverage, find in-network hospitals, or ask for a copy of their policy and get the PDF back in the chat.

Under the hood, a LangGraph agent routes each question. It decides whether the question is about the general conditions, which apply to everyone, or about the user’s specific policy. The general conditions were chunked, embedded with Gemini and indexed in Postgres with pgvector. Individual policies were processed and passed in as context. Both were cross-referenced with a structured database of hospitals and providers.

The system runs as several services. A Twilio client receives and sends messages. A backend runs the agent and the LLM calls. A database sits alongside, and side containers handle asynchronous work such as summarizing a conversation once it ends and fetching policy PDFs. Conversation memory lives in Redis; we built our own memory layer before frameworks offered one. Part of the job was designing the GCP architecture itself so the whole thing would hold up in production.

The bank’s fraud team was drowning. The existing rule set was so conservative that almost every account-to-account transfer got flagged, and analysts were checking them one by one.

We replaced the blanket rules with a learned score. An XGBoost model scores each transfer using the account’s transaction history, the characteristics of the transfer itself, and a set of engineered features. Choosing where to draw the line between false positives and false negatives was as much a business decision as a modeling one, and setting those criteria with the bank was a core part of the work.

It runs inside the bank’s own stack. The main pipeline is a batch job on the bank’s Databricks workspace that ingests transfers and writes scores back, so only the risky ones reach a human. We also deployed the same model as a Databricks serving endpoint, so it can score transfers in real time as they happen.

The old index was weighted by vote. IMCO’s team would get in a room and assign weights to each variable by consensus. The result wasn’t very scientific, and with so many indicators the subindexes ended up looking alike.

The new method separates judgment from arithmetic. Experts still decide which variables belong in each subindex, a question of meaning. Then PCA derives the weights from the data itself, so each subindex reflects how its indicators actually vary across states. The 2024 edition combines 50 indicators into six subindexes, from innovation and economy to law and government.

It became a recurring engagement. The methodology and code carried over from year to year; for each edition I helped with data collection, ran the pipeline, and adjusted the method as indicators changed. ICE 2024 is the edition below.

The question: does aircraft noise hurt how kids do in school? Mexico City’s flight routes changed, so some neighborhoods that used to be quiet started getting planes overhead, and others stopped. That shift is a natural experiment: compare schools’ standardized test scores before and after, depending on how their exposure changed.

The hard part was the data. The raw source was an enormous database of flight tracks, with a position (x, y, altitude) for every second of every flight: more than 17 million data points. Too big for one machine, it was downloaded into a sharded, distributed setup on AWS, where it was cleaned and transformed before any of it was usable.

From points to exposure. We laid a grid over the city and aggregated the track points into how many planes per hour crossed each cell, then matched cells to schools so exposure could enter the regressions alongside test scores.

The Air Filters and Student Learning project installed air quality monitors in 150 schools across Bogotá. Raw sensor readings don’t mean much to a principal or a city official, so I built reports and dashboards that summarize pollution by school and by locality.

The point was reproducibility: the team could regenerate every report and visualization with one run as new data arrived, instead of rebuilding them by hand. The reports went to the schools and to Bogotá’s city government, and the locality reports are public.

The study needed to know whether schools actually ran their air filters. The cheapest evidence is the filter itself: the more it’s used, the dirtier it gets. Field teams photographed each filter next to a white reference card.

The model finds two things in every photo. A YOLOv8 model, fine-tuned on images I labeled by hand, segments the reference card and the filter. Every photo used the same white card, so comparing the filter’s pixels to the card’s cancels out differences in lighting and camera, and the gap between them is the usage proxy.

Results go straight to the researchers. Scores land in a small Streamlit platform where the team can pull a report for any school, instead of eyeballing photos one at a time.

Matching people to jobs on a public portal. Mexico City’s Ministry of Labor (Secretaría del Trabajo) runs a job portal where residents post profiles and employers post openings. The goal was to suggest better matches between the two.

Before transformers, text meant bag-of-words. Profiles and job postings were represented with bag-of-words models plus engineered features, the standard toolkit at the time.

Built like production software, not a notebook. The system ran as a nightly batch: pull the latest data, rebuild features, and score matches. Data versioning and tracked experiments meant every model could be traced back to the data and settings that produced it.

Who the buscadoras are. Across Mexico, mothers and families of the disappeared have organized into collectives that search for their missing relatives themselves. No Están Solas (“You Are Not Alone”) documents that work, with interviews, a directory of contacts, and a data section.

The data section was ours. The project was led by Magic Grant recipients from the Brown Institute at Columbia, working with researchers at ITAM. I handled the data: cleaning INEGI’s public surveys, crossing them with private databases on the collectives, and turning the result into something people could explore.

Maps first. The centerpiece is an interactive Leaflet map of where collectives operate; click one to learn more. Other visualizations cross the collectives with crime data.