Headline

AI Engineer & Business Analyst | ML · LLMs · GenAI · AI Red Teaming | Data Analytics · Python · SQL · AWS

About

I got into AI because I wanted to understand how it fails. Not how to make it work — how it breaks. That curiosity landed me an 1-year role, red-teaming LLMs at AramAlgorithm.ai, where I designed adversarial attacks, stress-tested production models, and found vulnerabilities before they reached users. That experience changed how I build. Every system I ship now has a safety layer baked in from day one — not bolted on at the end. Since then I've built production ML and agentic AI systems: CortexAgent — a multi-agent RAG platform that answers SEC financial research questions using 4 specialized LLMs orchestrated via LangGraph. Built for institutional-grade financial analysis with 250ms latency and full audit trails. MigrationLens — an agentic code scanner that catches database migration failures before they happen. 100% recall on benchmarks, 0% false positives, 138 tests. Targets the 83% of migrations that fail in production. FairLend, ExperimentOS, ModelWatch — production ML systems covering credit scoring (EU AI Act compliant), Bayesian A/B testing, and real-time model drift detection. I sit at the intersection of AI engineering and business analytics. I care about systems that are explainable, safe, and actually useful to the people making decisions with them.

Location

Bengaluru

AI stack

codex

langgraph

langsmith

replit

Guardrails AI

Langflow

Work experience

AI Red-Team Engineer at Aram Algorithm, Mar 2025 - Feb 2026 · 1 yr, Washington, United States, • Designed and executed structured red-teaming frameworks to evaluate LLMs for safety, robustness, and misuse risks — covering prompt injection, jailbreaks, hallucination baiting, and adversarial manipulation across 3 model families. • Built automated pipelines to convert model evaluation outputs into structured, compliance-ready evidence artifacts — enabling engineering and compliance teams to act on findings without manual triage. • Translated AI safety requirements into measurable test plans and evaluation criteria in collaboration with engineering and compliance stakeholders — bridging the gap between safety research and production deployment.

Research Student at Vellore Institute of Technology, Jan 2025 - Apr 2025 · 4 mos, Chennai, AI-Driven DOOH & Retail Media Optimization system combining emotion recognition with personalized ad recommendation — bridging affective computing and business analytics. 1. Built a custom CNN model for real-time facial emotion recognition (happiness, sadness, anger, surprise) from grayscale images, integrated into a live ad targeting pipeline. 2. Engineered a hybrid XGBoost recommendation engine combining emotional state with purchase history, location, device type, and session behavior to serve personalized ad categories. 3. Designed cold-start handling via segment-aware profiling, ensuring relevant recommendations for new users with zero interaction history. 4. Delivered a fully modular Streamlit platform supporting real-time image input, CSV batch uploads, and interactive prediction dashboards — production-ready end-to-end.

Education

Integrated M.Tech — Computer Science & Engineering (Business Analytics), Computer Science and Business Analytics, Vellore Institute of Technology

12th Board of Secondary Education Of Andhra Pradesh, Sasi Educational Institutions

Public links

github: https://github.com/yaswankum2622-code

yaswanth-portfolio-beta.vercel.app: https://yaswanth-portfolio-beta.vercel.app/