featured research engineering
agent-economics-lab
Can an agent decision survive missing evidence, weak graders, and the cost of the work it creates?
This open-source, standard-library-only Python project audits whether the evidence around an agent supports a decision to expand it, supervise it, or switch it off. It combines trace-to-outcome evaluation, full-cost accounting, label-quality audits, and reproducible claims. It also publishes corrections when a result does not survive its own checks.
- 10
- public agent datasets audited under one protocol
- 31,389
- runs in the largest outcome-proxy audit
- 588
- controlled required-gate perturbations
These are dataset audits, controlled software experiments, and one small paired public case. They are not model rankings, production prevalence estimates, or deployment recommendations.
questions I am working on
- When outcome labels and cheap proxies disagree, which deployment claims still survive?
- Does letting a weak grader abstain improve it, or only change the population being measured?
- How should traces, outcome quality, human work, downstream loss, and missing evidence combine into a decision that can be audited?
work across the stack
- machine learning
- NLP, predictive modeling, ranking, monitoring, drift detection, and governance.
- generative AI
- Retrieval, synthesis, evaluation, guardrails, and observability for systems built on foundation models.
- agent systems
- Tool use, planning, routing, recovery, human handoffs, and the conditions under which a system should earn more autonomy.
technical notes and writing
- Found in the wild: reproducible audits of public agent datasets
- Dynamics blindness: your agent is more like a robot than a chatbot
- Cost per successful task, not cost per token
- Tool descriptions are infrastructure: on Anthropic acquiring Stainless
- The AI tools section, Davey Awards Under/Over Report 2026
external receipts
A few things the internet can verify without trusting this page.
- Institute of Analytics Fellow
- The Webby Awards Judging academy
- IEEE Women in Engineering Industry expert
- Semantic grouping with network graphs US Patent 11,748,453
- Navigating the Complexities of Generative AIs IBM Research · ADSA 2024
- Building Retrieval Augmented Generation UCLA Extension · Instructor
speaking
I occasionally speak about production machine learning, generative AI, agentic systems, evaluation, and responsible deployment.
Selected venues: PyData · Responsible AI Institute · IBM's Mixture of Experts podcast · IBM TechXchange (speaker and lab instructor)
For talks: gajjar.vyoma@gmail.com
teaching
Since 2021, I have taught machine learning, deep learning, production AI, RAG, generative AI evaluation, and responsible deployment through MIT Sloan (Great Learning), UCLA Extension, UT Dallas, and Maven.
about
I am a software-engineering-first applied AI engineer. I have spent the last decade building and operating systems from NLP and predictive models to RAG pipelines and tool-using agents.
The through line is production: how to make models measurable, observable, steerable, and useful once they leave the notebook.
I am interested in systems that learn from feedback without hiding failure, and in teams that value empirical questions, reproducible artifacts, and clear negative results.
San Francisco Bay Area.
Email LinkedIn GitHub