I'm Jerry Adams Franklin.
Senior AI Engineer @ Digital Currency Group · Deep Learning Engineer @ Intel · Data Scientist @ Nokia
>_
Building production-grade intelligent systems, from fine-tuning LLMs and orchestrating AI agents to shipping decentralized ML at scale.
// About
Building AI that decides, not just predicts.
I'm a Senior AI Engineer specializing in intelligent, decentralized AI systems that power real-world decisions. Most recently at Digital Currency Group, I architected production-ready AI agents with models like Qwen, LLaMA, Gemma, and Phi, orchestrated with LangChain, LangGraph, and Weaviate vector search, plus RAG, evaluation, and serving patterns used in production.
One of my proudest projects is VCScout-AI, an end-to-end autonomous agent that analyzed 300+ startup deals and directly drove $450K+ in AI-informed investments. My work spans federated learning, blockchain-integrated ML, and fine-tuning LLMs for high-stakes domains like financial trading and smart contract security.
I speak at conferences such as EthCC and NEARCON, guest-lecture at universities (including VIT-AP on production AI and LLM careers), and mentor at nonprofit hackathons like Opportunity Hack with Women in Computer Science at ASU, helping teams ship AI for real nonprofits in 24 hours.
Before DCG, I led TensorFlow CPU performance optimization at Intel, improving core operations significantly and earning the Division Achievement Award. I hold a Master's in Data Science from Northeastern University.
Welcome to my world. Grab a paddle; the AI waters are deep, and we're going swimming.
// Skills & Expertise
What I work with
Deep expertise across the ML lifecycle, from research and training to production deployment and monitoring.
My knowledge level in machine learning
Languages
Domain skills
LLMs & Generative AI
Qwen · LLaMA · Gemma · Phi · Fine-tuning · LoRA · QLoRA · PEFT · RLHF · vLLM
Agentic AI & RAG
LangChain · LangGraph · Weaviate · Vector DB · PydanticAI · Tool Use · ReAct
ML Infrastructure
GKE · H100 GPUs · ArgoCD · MLflow · Helm · Kubernetes · Docker
Deep Learning
TensorFlow · PyTorch · Computer Vision · NLP · Transformers
Federated & Decentralized ML
FLock Network · Federated Learning · Blockchain · Web3
Languages & Tools
Python · SQL · C/C++ · R · GCP · AWS · Azure · FastAPI · Git
// Experience
Where I've built things
Digital Currency Group- Architected and scaled decentralized AI systems on the FLock Network using Federated ML; forged strategic collaborations with Animoca Brands, Pundi AI, Chasm, and OneKey to build LLM-powered agents for VC analysis, financial trading, sports analytics, and smart contract security.
- Spearheaded fine-tuning of LLMs (Qwen, LLaMA, Gemma, Phi) and engineered a resilient ML pipeline using Argo CD, GKE (H100 GPUs), MLFlow, Helm, and Pub/Sub for scalable, production-grade training and inference.
- Designed and deployed VCScout-AI, a LangChain-based agentic RAG pipeline with Qwen-7B, Weaviate vector search, and custom evaluation frameworks to analyze 300+ startup deals, directly driving $450K+ in AI-informed investment decisions.
- Recognized as Top AI Arena performer; invited speaker at EthCC (Brussels) and NEARCON 2026 (San Francisco), including a lightning talk on training 7B LLMs on edge devices. Featured in the official FLock blog interview and tweet.
Foundry Digital- Spearheaded federated ML on the FLock network for decentralized models on chain; achieved top-performing status seven times across distinct tasks, with 1st place in two major ML hackathons and 2nd in another; mentored junior engineers on MLOps and model serving.
- Led a technical workshop at RIT on Machine Learning in Crypto, training and deploying production models on FLock and Akash Network.
- Featured in FLock interviews highlighting innovative work in the network.
OneBill Software- Led ML projects for Sales & Marketing, including Customer Account Clustering (K-Means) and a Product Recommender System (Collaborative Filtering) for 200+ accounts; implemented XGBoost churn prediction, retaining 93% of at-risk customers.
- Designed profitability, market opportunity, and pricing dashboards driving 20%+ quarterly financial gains; fine-tuned LLMs for industry-specific queries, boosting automated query resolution by 40%.
Intel Corporation- Optimized TensorFlow on Windows and pioneered an MLOps framework to streamline automation and acceleration of Neural Network benchmark performance on various Intel CPUs.
- Enhanced convolution operations inference latency by 400% using OneDNN; triaged and resolved critical TensorFlow CPU issues to production readiness.
- Built and released the TensorFlow binary to the worldwide TensorFlow community in collaboration with Google, achieving 10,000 downloads for TF 2.11 in a single day.
Intel Corporation- Built a telemetry data pipeline (Grafana/Graphite), developed an Isolation Forest anomaly detection system (92% accuracy), and created a BERT-based QA system for technical document retrieval.
Nokia- Designed Power BI dashboards, automated data pipelines (Flow), and contributed to go-to-market strategy solutions leading to a hyperscaler partnership.
Northeastern University- Taught Python, Machine Learning, and Data Analysis to 250 students; led practicum workshops on model development and evaluation; mentored student teams on applied ML projects.
OneBill Software- Developed SOAP/REST APIs and contributed to web service design using Postman and Maven.
// Education
Background
Dayananda Sagar University// Achievements
Recognition
Top Performer on the FLock Network
July 2024
Achieved top-performing model status 7 times across 7 distinct tasks on the blockchain-based FLock network. Featured internationally at EthCC Brussels and in the official FLock blog interview, tweet, and YouTube segment.
AIA Division Achievement Award, Intel
Nov 2022
Recognition for the TensorFlow Windows project, leading to Intel's ownership of the TensorFlow CPU build from Google and the first successful release for the Windows ecosystem.
// Speaking, mentoring & judging
Stages, classrooms, and hackathons
Speaking
AI Risk Summit & CISO Forum, Half Moon Bay
August 2026
Invited talk, "Federated Learning as AI Risk Infrastructure: The Aggregation Problem Enterprises Are Not Ready For," at SecurityWeek's AI Risk Summit & CISO Forum, Ritz-Carlton, Half Moon Bay. Selected from 300+ submissions. Covered the aggregation vulnerabilities unique to federated architectures and a governance framework for evaluating federated deployments as an enterprise AI risk strategy.
NEARCON 2026, San Francisco
Feb 2026
Lightning talk on training 7B LLMs on edge devices at the NEAR Foundation conference (Fort Mason), alongside speakers from PyTorch, Gensyn, Akash Network, and Hyperbolic Labs.
EthCC, Brussels
July 2024
Invited segment on decentralized AI and federated learning on the FLock network. How on-chain coordination and incentives change who can train models, what “production” means when compute is distributed, and lessons from shipping real federated workloads. Opens the official EthCC recording on YouTube (FLock segment).
Guest lecture, VIT-AP University
April 2026
Two-hour session for the School of Computer Science and Engineering on production AI: how LLMs are built and deployed, LoRA and QLoRA, why most models never leave the notebook, and a walkthrough of a production system from DCG.
Mentoring
Mentor, USAII Global AI Hackathon 2026
June 2026
Mentored graduate-track teams in a fully virtual, student-only AI innovation competition reaching 20,000+ students across 50+ countries. Guided teams on architecture decisions, evaluation strategies, governance frameworks, and lifecycle thinking for AI systems, pushing graduate-level reasoning beyond prototype-level demos. Hosted by USAII (United States Artificial Intelligence Institute).
Mentor, WiCS at ASU x Opportunity Hack
Spring 2026
Mentored university teams building real tools for nonprofits in 24 hours: AI architecture, LLM APIs, data modeling, and scoping so teams could ship. Hosted with Women in Computer Science at Arizona State University and Opportunity Hack.
Judging
Judge, University of Michigan AI Hackathon 2026
Oct 2026
Selected as a judge at the University of Michigan AI Hackathon 2026 (MHacks 26), the 20th installment of the Midwest's largest hackathon, with 1,000+ student builders, 521 participants, and $15,800 in prizes. Evaluated three-minute demos back to back on innovation, technical complexity, usability, and adherence to theme as teams shipped working prototypes in 24 hours. Sponsors included Amazon Web Services, Meta, Salesforce, D. E. Shaw & Co., and ElevenLabs.
Judge, IBM Bob 2.0 Hackathon, lablab.ai × IBM
Sep 2026
Selected by lablab.ai to serve as a judge for a 48-hour online competition with a $12,000 prize pool, hosted with IBM. Evaluated prototypes built with IBM Bob 2.0, an AI development partner that works with full repository context, to improve developer workflows such as onboarding, debugging, code review, and deployment. Assessed application of technology, presentation, business value, and originality.
Judge, UC Berkeley AI Hackathon 2026
June 2026
Selected to judge the UC Berkeley AI Hackathon 2026, one of the largest in-person AI/LLM hackathons globally, with 1,300+ participants and sponsorship from Anthropic, Microsoft, Adobe, Midjourney, and Amazon Annapurna Labs. Evaluated hacker projects on real-world application, technical execution, and responsible AI use across a full day of science-fair-style judging at the UC Berkeley Martin Luther King Jr. Student Union.
Judge, Web Data UNLOCKED Hackathon, lablab.ai × Bright Data
May 2026
Selected by lablab.ai to evaluate AI agent submissions at the Web Data UNLOCKED Hackathon, a $18,300-prize competition co-hosted by Bright Data. Assessed technical quality, LLM integration depth, and real-world viability of agentic AI systems built on live web data infrastructure.
// Publications & trade articles
Publications and trade articles
Publications
Published as Jerry Adams Franklin, Independent Researcher.
Partition Draws, Not Aggregators: Decomposing Result Variance in Federated LoRA Fine-Tuning Under Non-IID Splits
IEEE Access · Under review (submitted October 2026)
A pre-registered 204-run study that separates partition-draw variance from training-seed variance in federated LoRA fine-tuning (FedIT, FFA-LoRA, FLoRA at 1.1B and 3B scale). At Dirichlet alpha 0.1, partition draws and their interaction with the method account for nearly all held-out loss variance, and few-draw method rankings can reverse. The paper proposes the Partition-Draw Reporting Protocol. DOI · GitHub
A Measured Communication-Quality Frontier for Federated LoRA Fine-Tuning with Adaptive Phase-Switching
Scientific Reports · Under review (arXiv:2609.13512)
Measures per-round upload and download bytes for five federated LoRA protocols against held-out instruction-following loss and maps the resulting communication-quality frontier. The proposed ReverseAdaptive aggregator learns both LoRA factors, then freezes one once training-loss improvement plateaus, and sits at the frontier's knee: 40.5% less measured round-trip communication than FLoRA at a held-out loss cost of 0.0063. The same threshold transfers to LLaMA-3.2-3B without retuning, saving 30.0% communication. arXiv · DOI · GitHub
Fine-Tuning Large Language Models in Resource-Constrained Environments: Methods and Trade-offs
PeerJ Computer Science · Under review
This survey examines practical approaches to fine-tuning large language models under realistic hardware limitations. It covers parameter-efficient fine-tuning methods including LoRA and its variants, quantization techniques such as QLoRA and GPTQ, memory optimization strategies, and federated learning approaches for distributed training. The paper synthesizes empirical findings across these methods and proposes a decision framework for selecting appropriate configurations under varying resource constraints. Currently under peer review at PeerJ Computer Science (Q1).
Peer reviewer
ACM Computing Surveys
2026
Invited peer reviewer for ACM Computing Surveys, one of the highest-impact venues in computer science (Q1, Impact Factor 23.8), evaluating original survey manuscripts spanning machine learning, distributed systems, and AI infrastructure.
Journal of Supercomputing (Springer Nature)
May 2026
Invited peer reviewer selected by the Editor-in-Chief to evaluate original research manuscripts on high-performance computing, distributed AI systems, and machine learning infrastructure for one of Springer Nature's longest-running peer-reviewed computer science publications.
// Press
Press coverage
Coverage of my work in AI and federated learning.
// Featured
Highlighted work
VCScout-AI
End-to-end LangChain-based agentic pipeline leveraging Qwen-7B and Weaviate to autonomously analyze 300+ startup deals, directly informing $450K+ in investment decisions at DCG.
Decentralized AI on FLock Network
Pioneered federated learning models that achieved top-performer status across 4 different AI Arena cases. Featured at EthCC Brussels and in official FLock blog & interviews.
// Projects
Other work
Brain Tumor Detection
Developed a 12-layer TensorFlow CNN using Adam optimizer, Cross-Entropy loss, Dropout, and Batch Normalization, reaching 0.96 sensitivity for detecting tumors in MRI images.
Music Sentiment Analysis
NLP-powered system analyzing sentiment in music lyrics using transformer-based models and custom fine-tuning pipelines.
Recommender Systems (Amazon Reviews)
Built a collaborative filtering system using PCA-reduced user-item matrix, Agglomerative Hierarchical Clustering, and Matrix Factorization, achieving MAE of 0.096.
Imbalanced Datasets: Preeclampsia Prediction
Case study on the PROTECT health database: implemented Bootstrapping, SMOTE, and Semi-Supervised Learning with Autoencoders and classifiers to achieve 96% accuracy in predicting Preeclampsia in pregnant women.
Diabetes Prediction in R
Trained Logistic Regression, SVM, Random Forest, and XGBoost models; achieved 0.92 sensitivity; handled sparse/imbalanced data using ROSE sampling and Boruta feature selection.
// Contact
Let's connect
Feel free to reach out to learn more about me and the work I do. Always happy to chat about AI, engineering, or new opportunities.
Fairport, NY 14450 · franklin.je@northeastern.edu