WorkCase study
Trust analytics for AI agents
When software agents act on their own, how do you tell from their behaviour whether they can still be trusted?
- Role
- AI Engineering Intern
- Organisation
- Infosys Limited
- Dates
- Feb 2026 – Aug 2026
- Location
- Bangalore, India
- Tags
The question
During my internship at Infosys Limited I joined a team building a framework for governing AI agents. Governance rules say what an agent may do. They do not say whether an agent is still behaving as it should. My part was the analytics layer that tries to answer that second question from observed behaviour.
What I did
I worked on the trust-analytics layer. In plain terms, that meant:
- turning raw agent activity into features that describe behaviour;
- detecting adversarial behaviour with both supervised and unsupervised methods;
- monitoring drift in those features with the population stability index, so that a change in what agents do would be noticed;
- building dashboards in Streamlit and Dash so that people other than me could read the results.
I also built an adversarial simulation we called Attack Day. It uses synthetic attacker agents to stress-test how the framework scores trust. Separately, I co-authored the team's Project Fest submission and presented it through its review rounds.
What failed and its limits
My attack data was synthetic. I wrote the attacker profiles myself, so my detectors could learn my own definition of an attack, and a good score against them says little about real adversaries. The models I used were standard ones; I did not calibrate them or report their uncertainty, so a score could look more precise than the evidence behind it. Everything I measured, I measured inside the simulation.
What I would do now
I would treat trust as sequential Bayesian inference. Each observation of an agent updates a belief about whether it is reliable, and the belief carries its own uncertainty. A change in regime, such as an agent that starts behaving differently, becomes something the model can detect and be tested on, against adversaries fixed in advance rather than invented along the way.
That is the same kind of problem as inferring a hidden state from noisy choices, which is the subject of my research note. The internship is where I met the question, and the research is where I want to answer it properly.
Evidence
Built the trust-analytics layer of a team-built Agent Governance Framework: feature engineering, supervised and unsupervised adversarial detection, and PSI drift monitoring, with Streamlit and Dash dashboards
Built 'Attack Day', an adversarial simulation with synthetic attacker agents that stress-tests the framework's trust scoring
Co-authored the InStep Project Fest 2026 submission with a team of 4 and led the presentation through 3 review stages, including a session with Infosys founder N. R. Narayana Murthy and CTO Rafee Tarafdar
Next · Research note
Strategy or state?
How much of the predictive advantage of tiny recurrent networks over classical reinforcement-learning models comes from hidden engagement states, and how much from different learning rules?
Got a question about it?
leyva.