Machine Learning Engineer
I build production ML systems — from real-time inference at Apple Sports to LLM reasoning pipelines (GRPO, SFT, tool-use RL) and high-throughput risk engines at Credit Suisse. 10+ years shipping ML across fintech and big tech.
Career progression across fintech and big tech
May 2022 — Present
Leading ML infrastructure initiatives for Apple platforms. Large-scale model training, optimization, and on-device inference. NDA-bound work.
Researching and deploying ML pipelines for Apple platforms. Large-scale model training, optimization, and on-device inference. NDA-bound work.
Prototyped NLG architectures and text processing utilities for on-device intelligence features.
2011 — 2022
Redesigned multithreaded Credit Risk Calculation Engine for Wrong Way Risk trades. Built secure API server with Kafka/Hazelcast cluster monitoring and failure detection. Created Issuer Linking methodology improving counterparty data quality.
Designed quantitative model benchmarking credit risk (loans vs bonds). Implemented WWR Calculation Framework (FINMA/PRA compliant). Developed RWA simulation services for stress testing counterparty portfolios.
Built SMEBot — credit risk chat assistant using NLP/DL. Intent classification with Word2Vec, GloVe, BERT ensembles. QANet for extractive QA from financial PDFs. Used by 200+ analysts.
Co-developed Digital Regulatory Reporting (DRR) framework accepted by FCA (UK). Automated SQL generation and testing databases for Wealth Management.
Academic foundation and research experience
Georgia Institute of Technology
2021 — 2022
Veermata Jijabai Technological Institute (VJTI)
2013 — 2017
Emory University — Clifford Lab, Biomedical Informatics
2021 — 2022
Designed and deployed audio signal processing applications using React, Python, and shell automation for biomedical data analysis. Published work on indoor localization using sparse camera networks and BLE trilateration.
VJTI — Center of Excellence in Nonlinear Dynamical Systems
2015 — 2017
Researched convex optimization algorithms for degaussing systems using MATLAB and OPERA electromagnetics software. Published 4 papers on hexagonal coil systems for naval demagnetization.
Selected ML, LLM & agent projects
Reproduced DeepSeek-R1 reasoning pipeline on Qwen2.5-3B-Instruct. Three-stage training: distilled SFT → GRPO with verifiable rewards → rejection sampling. Ablation study across GSM8K (42% → 59%), MATH, and reasoning benchmarks. Full training infra with Ray Train, vLLM, TRL.
Trained Qwen2.5-3B-Instruct to learn its own retrieval policy via self-play rollouts with rejection sampling + GRPO optimization. Agent decides when to search, what to query, when to stop. HotpotQA F1: 31% → 47%, retrieval calls reduced 4.2 → 2.8/question.
Built a local-first coding agent harness with sandboxed execution, file system tools, test running, and git integration. Implements planning→coding→verification loop with multi-agent orchestration (planner, coder, reviewer). Supports LangGraph/LangChain, evaluates on SWE-bench, HumanEval.
Automated evaluation suite for instruction-following, hallucination detection, and safety. Implements LLM-as-judge, self-consistency checks, and adversarial prompt generation. CI/CD integration for regression testing model updates.
Let's build something together
Google Scholar
