Abraham Guan

MACHINE LEARNING

I've worked on benchmarking foundation models, agent systems, and retrieval architectures across multiple industries, including manufacturing, biology, and finance.

+4 Internships
#1 Public
university
+12 Projects

Training and benchmarking foundation models, reinforcement learning, RAG, LLM as a judge

Python, PyTorch, SQL, Docker, AWS, LangChain

PROFESSIONAL EXPERIENCES

NASA Ames Research Center
Jan – Aug 2026 8 months

Machine Learning Research Intern · Remote

Took spaceflight-signal detection from 0.5 AUROC (chance) to 0.83–0.97 across five tissue types by tracing the old null result to an evaluation flaw, not biology. Built the team's benchmarking harness over 2.3M unified samples.
Aerospace research · Space biology

Problem

The standing conclusion was that spaceflight leaves no detectable signal in tissue: detection sat at 0.5 AUROC, no better than chance. The original random train/test splits also leaked batch effects, so models were scoring which lab produced a sample rather than whether it flew.

Approach

Traced the null result to an evaluation flaw rather than a biological one, and rebuilt the protocol on study-grouped cross-validation so every model is judged on biology. Built the team's reproducible benchmarking harness, the first fair comparison of foundation models on space biology data, running five foundation models against classical baselines across three tasks on ACCESS GPU clusters, over a 6-stage preprocessing pipeline that unified 2.3M samples with NASA spaceflight RNA-seq into one comparable gene space.

Impact

Detection rose from 0.5 to 0.83–0.97 AUROC across five tissue types, overturning the prior conclusion.

Foundation models

Benchmarking

Study-grouped CV

Batch-effect diagnosis

RNA-seq

GPU clusters

Python
Molex
May – Aug 2026 4 months

Machine Learning Engineer Intern · Fremont, CA

Cut mirror-swapping time 74% with a LambdaMART ranking model, and a custom Gaussian Process kernel that dropped prediction error 30% below the team's hand-weighted baseline. Owned 14 models end to end.
Electronics manufacturing · Optical connectors

Problem

Choosing the best performing parts meant slow manual trial and error. The team's hand-weighted model left days of testing on every new optical module, and the underlying data was scattered across Excel files.

Approach

Trained a LambdaMART ranking model, evaluated with nDCG@k, to predict the best performing parts. Built a custom Gaussian Process kernel that captures a repeating pattern in each optical module. Owned 14 prediction models and the pipeline end to end, from scattered spreadsheets to one clean dataset, through feature engineering with ANOVA, correlation and clustering studies, to benchmarking nonlinear methods against simple baselines.

Impact

Mirror swapping time down 74%, prediction error down 30% against the hand-weighted model, and days of testing removed from every new module.

LambdaMART

Learning to rank

nDCG@k

Gaussian Processes

ANOVA

Clustering

Feature engineering
Umbo Energy
Sep – Dec 2025 4 months

Machine Learning Engineer Intern · Remote

Designed an offline RL system (behaviour cloning, surrogate model, CQL) for a 7% cut in annual energy use, with conservative Q-learning enforcing safety during exploration.
Energy · Building systems

Problem

Cutting energy use meant searching over control policies, but exploring on live equipment risks damaging it, so the policy had to be learned offline from logged data rather than by trial on the real system.

Approach

Designed an offline reinforcement learning system on a three-model architecture: behaviour cloning, a surrogate model, and CQL. Implemented Conservative Q-Learning to enforce safety constraints and keep exploration away from actions that would damage equipment.

Impact

A 7% reduction in annual energy consumption, with safety constraints holding throughout exploration.

Offline RL

Conservative Q-Learning

Behaviour cloning

Surrogate modelling

Safety constraints
Stockbit
Jul – Sep 2024 3 months

AI Engineer Intern · Jakarta, Indonesia

Built the retrieval layer for an investing RAG assistant (BM25 plus a ticker-and-metric query parser) at 0.87 Recall@100, with an evaluation framework that separates retrieval failure from generation failure.
Fintech · Retail investing platform

Problem

An investing assistant has to resolve stock tickers and financial metric synonyms before it can retrieve anything useful, and when an answer came back wrong there was no way to tell whether retrieval or generation had failed. Research on the platform took over five minutes.

Approach

Built the retrieval layer, pairing BM25 keyword search with a query parser that resolves tickers and metric synonyms to canonical fields. Designed a staged evaluation framework that separates retrieval failure from generation failure: first whether the chunk was retrieved at all, then answer faithfulness scored with LLM-as-a-Judge validated against human labels. Engineered a BigQuery and Kubernetes pipeline for real-time financial data, and fine-tuned Llama 3 for a customer FAQ assistant on the same stack.

Impact

0.87 Recall@100 and 0.89 F1 on a held-out query set, and research time cut from over five minutes to under 30 seconds.

RAG

BM25

Query parsing

LLM-as-a-Judge

BigQuery

Kubernetes

Llama 3 fine-tuning