Mithin Sagar
Skip to content
05Research

Papers, and thework behind them.

One conference paper presented, two research frameworks with reproducible evaluations. Everything here has a repository; nothing here claims a number it cannot recompute.
Publications
1 presented
Frameworks
2
Focus
Explainability & robustness
Publications

The archive.

Click a record to open its abstract.
Abstract

Idle cloud resources sit unused and forgotten, quietly driving up cost and widening the attack surface. This work presents an event-driven Boto3 system that scans EC2, EBS, S3, IAM and RDS for waste, then reclaims it on a scheduled Lambda cron with dry-run simulation and protected-tag exclusions guarding against accidental deletes. A supervised classifier trained on utilisation metrics predicts a keep, review or delete recommendation for each resource, and statistical outlier detection surfaces anomalous cost spikes alongside the rule engine.

Authors
Mithin Sagar S
Status
Presented · publication pending
DOI
[PENDING]
  • Cloud cost optimisation
  • AWS
  • Resource lifecycle
  • Security posture
  • Applied machine learning
Implementation on GitHub
Research work

Frameworks,not findings.

Two bodies of work that behave like research even though they live in a repository: a formal metric, a controlled comparison, and a result that can be re-run.
Technical report2026

XAI Attack & Defense Framework with Few-Shot Learning

Quantifies how far explanations can be manipulated without touching a prediction, using a formal Explanation Drift metric across four attack families and three security datasets — then proposes four defenses, the strongest cutting drift by up to 91%.

91%drift reductionCase study →Source ↗

Collaborators on the original technical report: Gokul Ram K, Kishore A G.

Framework & evaluation2026

EXAI-ResumeIntel — Explainable Resume Analysis

Exact Shapley values satisfying all four game-theoretic axioms, independently validated by LIME across 300 perturbation samples with 100% directional agreement, evaluated on 2,484 resumes across 24 job categories.

70.73%overall accuracyCase study →Source ↗

What I keep
coming back to.

    01

    Explainability that survives contact

    An explanation is only useful if it holds up when someone tries to break it. I care about measuring that, not asserting it.

    02

    Calibration and honest uncertainty

    A confidently wrong answer is worse than no answer. Reported confidence should match observed accuracy.

    03

    Retrieval over recall

    Systems that fetch evidence and cite it, rather than systems that remember and hope.

    04

    Agentic pipelines with owned stages

    One agent, one responsibility, one auditable handoff — so a failure has an address.

See how it shows up in the work
Collaborate

Let’s build somethingmeaningful.

Open to internships and AI/ML roles, research collaborations, or a conversation about something you are trying to make work. I reply to everything.

EmailGitHubLinkedInInstagramRésumé