Aman Singh

Aman Singh

MS CE – Arizona State University

"How do I solve a hard problem?"

Insight — when complexity suddenly makes sense.

— Aman Singh

💡 Do You I Tech?

I do — and that's why I spend my time building innovative projects. Hi, I'm Aman, an MS in Computer Engineering student at ASU. I've worked on projects across AI, ML, Reliability, Programming Languages, Built Servers .

Work Experience

Projects

OOD vs ADV thumbnail

OOD vs ADV

Problem —

Out-of-distribution (OOD) inputs and adversarial (ADV) attacks both trigger detectors (e.g., Mahalanobis). But their root causes and mitigations are different — how do we reliably tell them apart so systems choose the right defense?

What I did —
  • Designed a simple, post-hoc solution
  • Separated OOD vs adversarial inputs after detection
  • Enabled targeted, low-overhead mitigations
OOD vs ADV vs ID

OOD / ADV vs ID

Problem —

OOD inputs and adversarial attacks are both flagged as anomalies, but they require different defenses.

What I did —
  • Designed a simple, post-hoc solution
  • Separated OOD vs adversarial inputs after detection
  • Enabled targeted, low-overhead mitigations
Silent Data Corruption

Silent Data Corruption (SDC)

Problem —

SDC can silently corrupt model data and outputs, reducing reliability. Many defenses lack reproducible reference implementations.

What I did —
  • Implemented reproducible SDC defenses and injection tools
  • Built test harnesses for large-scale benchmarking
  • Released artifacts for reproducible research
LLM reasoning

LLM Reasoning — Puzzles & Failure Modes

Problem —

Large language models often fail on structured logical puzzles and reasoning chains, producing plausible but incorrect outputs.

What I did —
  • Designed puzzle-based benchmarks
  • Analyzed stepwise traces to build error taxonomies
  • Proposed interventions to reduce systematic errors
Information retrieval

Information Retrieval

Problem —

Large SLMs are compute-heavy; smaller models need augmented retrieval to match retrieval quality of huge models.

What I did —
  • Built memory-augmented retrieval to improve small-model performance
  • Reduced compute while maintaining retrieval quality
  • Open-sourced code and experiments
Functional correctness

Functional Correctness — LLM Agents

Problem —

LLM agents can execute tasks incorrectly due to reasoning/execution gaps; few systematic taxonomies exist to guide fixes.

What I did —
  • Analyzed agent executions and built an error taxonomy
  • Proposed verification loops and structured feedback
  • Included example traces and remediation patterns