IAM-CausalKit
LiveDraw the DAG, get a formally verified adjustment set and the R/Python code.
d-separationPerkovićDAGitty I/O
AI × Epidemiology
I am a PhD epidemiologist who writes the software — so I can build the system, and tell you whether its answer is actually right.
{Age, SES}.Not whether it sounds right — whether it is epidemiologically, clinically and methodologically correct.
Nothing that decides a number lives in a UI component, and the tests run inside the build.
By day. Which is why the tools assume messy data, thin staffing and a decision due this week.
In production
Every link opens something live, or source you can read.
Draw the DAG, get a formally verified adjustment set and the R/Python code.
Computational epidemiology: causal DAGs, scientific RAG over PubMed, SEIR and renal digital twins.
A methodology engine that carries doctoral research from PICO to GRADE. Not a chat wrapper.
AI-assisted property portal. An agent publishes from the street with a voice note — no laptop.
Heterogeneous investments on one basis. Risk read like a meta-analysis: P5–P95 on a forest plot.
Essential Public Health Functions assessment for state and district health authorities.
FHIR R4 client, clinical terminologies, PHI detection and masking, multi-provider LLM layer.
Measles care pathway for first-level health workers. One offline file — it works where the signal does not.
Not a chatbot — a virtual office. NEXUS coordinates a team of specialised AI workers, each with its own model, role and memory. You delegate from Telegram; they execute and report back.
Incubator
Each one carries the limitation a reviewer would ask about first.
Pap smear classification with transfer learning and Grad-CAM explainability.
Synthetic data only. Ethics approval pending. Not a diagnostic device.
Decision support for tuberculosis preventive therapy at the first level of care.
Under submission to a PAHO/TDR grant. Not funded, not deployed.
Director’s module bank specialized for a cervical cytology research line.
One institution. Not generalized to other research lines.
Send a photo by Telegram, get back a full brand package: reel, card, story and copy.
Working MVP on Telegram. WhatsApp blocked on Meta API approval.
The multi-city generalization of KzaHub, one deployment per city.
Phase 1 implemented. Not yet tested end to end.
Governance for a family enterprise: score initiatives, assign owners, hold the weekly review.
Phase 0: offline in the browser. Multi-user version specified, not built.
01
Thirty minutes. If AI is the wrong instrument, I say so.
02
Deliverables, timeline, fixed price, stated assumptions.
03
Code, tests and rationale — so your team maintains it without me.
Insights
Hay un error de ajuste peor que bloquear un efecto: crear uno que no existía. Y los datos no te van a avisar.
Four things that decide the deal, none of which are the model — written from the side of the table that has to sign and answer for it.
The connection drops. The patient is still in front of you. That sentence was the whole specification.
About
For a decade I did the serious version of this work: a PhD in epidemiology, National Director of Epidemiology for a ministry of health, project coordinator with PAHO/WHO, graduate faculty in research methodology.
And I kept hitting the same wall: the methods exist, the papers exist — what doesn’t exist is someone who turns rigorous science into working software. So I founded IAM-Studio, an AI studio born inside public health rather than imported from tech. What we don’t build is a generic chatbot with a health sticker on it.
Also from the studio
The same engineering discipline, applied outside health. This line funds the research.
WhatsApp and web assistants that answer, book and sell around the clock.
Leads, follow-up, CRM and reporting wired with n8n.
Visual identity, premium content, and FORMA·3D — 3D scanning printed and painted by hand.
Contact