Selected work

Things (hopefully) worth showing.

I'm too cynical to say that I'm proud of anything, but here are some things that I have spent significant time on. The current one isn't public yet, so it appears in outline only; everything else links to the real thing.

building now

current project · not public yet

Trust infrastructure for AI agents

Agents take actions. You take responsibility. Observability platforms tell you what happened, after it happened to a real customer. I'm working one step earlier: run the agent in a world you control, high-fidelity simulations of the platforms it operates, and judge it by the side effects it actually causes rather than what it reports. The adversarial states that define a trust boundary almost never show up in live traffic; in a controlled world you can manufacture them on demand.

controlled simulation worlds · deterministic capture & replay · side-effect-level evals · trust envelopes made explicit

Email me about it No public link yet. More when it's ready.

shipped · still live

co-founder & technical lead · oct 2025 – june 2026

Simply-Useful

Task management for ground-level teams: dispatchers, coordinators, and operators. Create a task by voice or text, assign it to anyone (even people who don't have the app), and track status as updates come back by SMS or email. I own the whole system, from the mobile client to the backend.

React Native (iOS + Android) · Django on DigitalOcean · in-app AI · two-way Google sync · CI/CD, analytics, error monitoring

App Store Google Play Web app simply-useful.com →

open source · research

Code and datasets behind papers I co-authored at Microsoft, released so others can build on them. All on github.com/meniData1.

EMNLP 2024

Fine-Tuning or Retrieval?

Data and evaluation code from our study comparing unsupervised fine-tuning against retrieval-augmented generation for injecting knowledge into LLMs, including facts a model never saw in pre-training.

GitHub arXiv ACL Anthology

2025

Knowledge-Instruct

Datasets and scripts for continual pre-training from limited data using instructions: a cheaper way to teach a model new, long-tail knowledge without forgetting what it already knows.

GitHub arXiv

2024

The Cocktail Effect

Training code for our ablation study on how multi-task fine-tuning "cocktails" change LLM performance in domain-specific settings like finance, across dozens of task combinations.

GitHub arXiv

2024 · dataset

ReMatch: MIMIC to OMOP

A gold-standard schema mapping between the MIMIC-III and OMOP medical data models, released as a benchmark for retrieval-enhanced schema matching with LLMs in healthcare.

GitHub arXiv