Decentralized AI · Energy Forecasting · ML
BittBridge · UCONN · Project Manager Lead · Spring 2026
Subnet 183
Building a decentralized AI energy forecasting platform from the ground up
PM leadership
Decentralized AI
Architecture design
Energy forecasting
ML modeling
Stakeholder management
Bittensor
Cross-functional leadership
Challenge
What problem were we solving?
BittBridge is a student-led decentralized AI initiative at UCONN. The core infrastructure already existed — but there was no real, live problem for students to solve on it. The team set out to build Subnet 183: a brand new energy forecasting subnet designed from scratch, where students would act as miners competing to predict real-time electricity demand for the New England grid.
The challenge was not just technical. We had to design a live production problem statement that 100+ students with varying technical backgrounds could meaningfully participate in — with real data, real scoring, and a real leaderboard updating every few minutes.
Complexity
What made it difficult?
The team brought together strong expertise across ML, data systems, and engineering — but everyone started with a different mental model of how the miner-validator architecture would actually work end to end. Without a shared picture of the full pipeline, coordination was difficult and decisions were getting made in isolation.
The hardest part was not the technology — it was alignment. Different team members had different assumptions about who owned which step, how data would flow, and what students would actually need to do to participate.
I created the workflow diagram that mapped the entire end-to-end process — from how students would clone the repo, to how predictions would flow through the validator, to how scores would be stored and displayed on the leaderboard. This became the shared reference that brought the team onto the same page.
Miner-Validator process flow — created to align the full team on how Subnet 183 would work end to end
Process
What did I work on?
My role as PM Lead spanned research, coordination, architecture design, communication, and evangelism across the full lifecycle of Subnet 183.
1
Architecture research and workflow design
Researched the Bittensor miner-validator architecture deeply to understand how to adapt it for an energy forecasting problem. Designed the end-to-end workflow diagram that documented every step from student onboarding to live scoring.
2
Team coordination and weekly planning
Ran weekly calls across the cross-functional team — ML lead, data systems, engineering — to keep alignment, surface blockers, and drive decisions. The workflow document became the standing reference for every meeting.
3
Demo presentation and hackathon data pipeline
Prepared the official BittBridge demo deck used across all university presentations and the hackathon event. Also contributed to documenting the hackathon data pipeline — how weather data, historical load, and live ISO-NE grid data flowed through the system.
4
University events and promotion
Presented BittBridge at two UCONN events — Data Disrupted: Sports Leadership & AI in Action, and AI Day Hartford — walking attendees through what decentralized AI is and how Subnet 183 works. Managed the booth, answered questions, and represented the project publicly.
Hackathon data pipeline — showing how weather data, historical load, and live grid data flowed through the system to score miners every 5 minutes
Turning Point
When did it come together?
The moment Subnet 183 went from a team project to something with real public presence was at the Data Disrupted event at UCONN. People stopped at the table, engaged with the demo, picked up printouts, and asked genuine questions about decentralized AI and what students were building. It was the first time the platform had an audience outside the team.
Presenting something this technical to a non-technical audience — faculty, industry guests, university leadership — reinforced how important the PM role was. Not just building the product, but being able to explain it clearly to anyone who walked up.
Data Disrupted: Sports Leadership & AI in Action — UCONN
Presenting Subnet 183 to attendees at the UCONN AI event
BittBridge team — Project Manager Lead role
Personal Achievement
From building the platform to competing on it
Once Subnet 183 was live, I participated as a miner myself — submitting my own energy demand forecasts to the same platform I had helped build. This meant building a forecasting model from scratch, iterating through multiple approaches, and competing against 75 other student miners in real time.
I went through four model iterations before finding what worked:
1
Moving average baseline
R² below 0.50 — confirmed that smoothing past values alone was nowhere near enough to forecast demand six hours ahead.
2
CART model with 66 features
Training R² of 0.998 but validation dropped to 0.71 — a textbook overfitting case. A single decision tree, no matter how feature-rich, memorizes rather than generalizes.
3
LSTM deep learning model
Multiple rounds of tuning — adding load lag features, enabling input standardization, extending lookback to 3 hours of history. Final validation R² of 0.86 and MAPE of 5%. Forecasts were still too flat and missed sharp daily swings.
4
HistGradientBoosting — final model
Trained on the full 244,000-row dataset with all 66 features. Validation R² of 0.925 and MAPE of 3.66% — outperforming the LSTM on every metric. The key insight: since load lag features already encoded temporal history explicitly, the LSTM's sequential memory wasn't adding much.
UID 2 (my miner) tracking actual ISO-NE energy demand — 7-day time series. White line = actual demand, blue line = my forecast.
Average Miner Performance Leaderboard — UID 2 ranked #1 out of 76 miners over a 7-day window (scores are dynamic and shift over time)
Reaching Rank 1 wasn't just a personal milestone — it validated that understanding a platform deeply enough to build it also means understanding it deeply enough to win on it. The same curiosity that drove the PM work drove the model work.
Impact
What did Subnet 183 unlock?
100+
Student miners on the live platform
#1
Ranked out of 76 miners over 7 days
2
University AI events presented at
Subnet 183 became the first live decentralized AI energy forecasting platform built and run by students at UCONN — giving 100+ participants hands-on experience with real ML workflows, live data pipelines, and competitive model evaluation. Beyond the numbers, it demonstrated that decentralized AI is not just a concept — it is something students can build, run, and participate in meaningfully.