CV

Education

• MS in Applied Statistics, New York University, Aug 2024 – Expected Aug 2026
  • Specialization in Machine Learning
  • Coursework: NLP, Deep Learning, Data Engineering, Distributed Systems, Big Data, Bayesian Methods, Biostatistics
• BS in Statistics, Baylor University, Aug 2020 – Aug 2024
  • Minor in Biology · Honors College · BU Chess Club
  • Coursework: Data Structures & Algorithms, Linear Algebra, Calculus, Database Systems, System Design, Operating Systems

Work Experience

Machine Learning Engineer — NYU AI Whistleblower Initiative | CORDA AI Fellow Apr 2026 – Present · Remote

• Architected an RLHF pipeline over a large language model, collecting annotations from 10+ experts and converting pairwise ratings into a reward model, improving precision by 30% over baseline zero-shot prompting
• Built a drift detection layer computing KL-divergence across consecutive model output distributions, triggering remediation workflows when shift exceeded a tuned threshold across 500+ weekly sessions
• Constructed a semantic clustering pipeline vectorizing annotated session transcripts with sentence-transformers and FAISS, reducing retrieval latency by 60%
• Evaluated prompt engineering methods on expert-labeled datasets; selected chain-of-thought, reducing hallucinated legal citations by 22% on held-out jurisdiction-specific tests
• Feedback fine-tuned via LoRA adapter training, cutting full model retraining cost by 80% while maintaining within 2% on domain evals

Machine Learning Engineer — Supervised Program for Alignment Research (SPAR) Feb 2026 – May 2026 · Remote

• Investigated AI safety properties of frontier LLMs using Bayesian methods to model AI consciousness; designed experiments across 3 models using Python and PyTorch
• Fine-tuned LLMs to generate 6 semantically equivalent question variants per safety indicator; measured cross-variant output distributions to stress-test alignment stability, reducing variance by 30%
• Developed mathematical models characterizing ground-truth drift in LLM judgment across model generations; analyzed how scaling and RLHF-driven training affect model behavior
• Built statistical analysis over response distributions including Likert histograms, cross-model EDA, and semantic analysis of chain-of-thought justifications across 50+ safety-relevant indicators
• Open-sourced full evaluation framework on GitHub for reproducible AI safety research

Software Engineer, AI/ML — Replays AI Sep 2025 – Apr 2026 · New York, NY

• Designed ingestion pipeline writing 5M+ structured play-by-play events into PostgreSQL as system of record; stored raw video in GCP referenced by pointer, cutting compute costs by 40%
• Preprocessed using Python to handle schema validation, player disambiguation, and deduplication; anchored video segmentation to event timestamps before Computer Vision inference
• Orchestrated 3 parallel agentic inference pipelines spanning event feature extraction, CV play classification, and task-split LLM summarization; reduced inference latency by 35%
• Decoupled storage and inference from the frontend using REST APIs; implemented recap caching sustaining sub-second response times
• Built React and TypeScript mobile app rendering live rankings, game recaps, and CV-scored highlight reels; reduced mean load time by 28% via component memoization

Founder & ML Engineer — Synthure | Care Clarity AI Jul 2023 – Jul 2025 · New York, NY

• Built RAG retrieval pipeline over Snowflake for medical code classification and claim denial routing; reduced policy misinterpretation errors by 94% across production clinical workflows
• Engineered typed intermediate representation and data quality gate validating schema, deduplication, and entity extraction confidence before downstream inference
• Designed multi-agent orchestration with task-split LLM routing using Claude APIs; directed high-volume entity tagging to lighter models and plain-language generation to frontier LLMs
• Implemented gated controlled generation constraining outputs to source-grounded entity rewrites; reduced hallucinated content across 60K+ records
• Deployed vLLM inference at under 1.8s p95 latency on AWS within MongoDB and Redis microservices; enforced HIPAA-aligned JWT access control and input validation across all services

Projects

CSCI 2271 Computer Vision — GitHub

• Trained diffusion world model on 737K+ frames with VAE, CNN reward model, and PPO agents

RobustSight: Advancing AI Safety and Alignment — GitHub

• AI safety framework investigating the intersection of adversarial robustness and interpretability

Publications

Agentic AI Systems for Clinical Reasoning

Kannappan, A. et al. (2026). "Agentic AI Systems for Clinical Reasoning." Book Chapter.

Multiple Myeloma Gene Signatures

Kannappan, A. et al. (2024). "Multiple Myeloma Gene Signatures." The Oncologist.

Skills

Languages: Python, TypeScript, SQL, R, C++

AI/ML Frameworks: LangChain, LangGraph, PyTorch, HuggingFace, vLLM, Scikit-learn, TensorFlow, FAISS, RAG, RLHF, Prompt Engineering, Computer Vision

Frontend: React, REST APIs

Cloud & DevOps: AWS (S3, EC2, Bedrock), GCP, Docker, Kubernetes, CI/CD, Git, PostgreSQL, MongoDB, Redis, Snowflake