BittBridge AI: UConn students building decentralized AI in the real world.
The BittBridge team delivers hands-on decentralized AI systems development, tackles real business challenges, and translates research into production-ready infrastructure with support from the Yuma partnership.
Overview
UConn is the first academic institutions offering experiential learning in decentralized AI on Bittensor. Through the BittBridge initiative and partnership with Yuma, graduate students receive onboarding, technical mentorship, and operational support while building real systems.
- Live leaderboard backed by W&B validator runs
- Hands-on decentralized AI systems development at UConn
- Industry-supported mentorship model through partnership with Yuma
- Production-ready infrastructure for custom ML workflows
Why decentralized AI
- Democratization: Distributed validators and miners improve uptime and operational continuity.
- Transparency: On-chain and auditable workflows make decisions easier to inspect and trust.
- Access: Open participation enables universities, builders, and teams to contribute directly.
- Incentive: Network-based incentives reward useful performance and accelerate experimentation.
Operating scope
- Bittbridge subnet: Decentralized energy load prediction in New England workflow using ISO-NE system load data.
- Challenge-response design: Validators query miners every 5 minutes.
- Delayed evaluation: Scores computed after a 10-minute truth window.
- Pre-built Models: Moving Average + Linear Regression + CART + RNN + LSTM.
How it works on-chain
Watch one full round flow through the subnet — each step pulses through the network on loop.
01
/ 06
Challenge dispatch
Validators broadcast a load-forecast challenge to every miner in the subnet.
Average Miner Performance Leaderboard
Each miner's mean moving average score over the selected window. The delta (▲/▼) shows the difference between the current score and the window average (= current score − window average), indicating whether their current score is higher or lower than their recent average.
⏳ Calculating average performance...
Live surface
Miner performance leaderboard
The dashboard pulls validator telemetry directly from W&B and ranks miners using the latest moving-average scores published by active runs, giving UConn students a live view of how the subnet behaves during hackathon work.
Waiting for W&B data...
Miner Forecast vs Actual Energy Demand: Time Series Plot
Before Apr 26, 2026, 6:00 PM ET, forecasts used a 5-minute horizon, so miner lines use the log time on the x-axis. From that time onward, forecasts target six hours ahead, so miner lines use log time + 6 hours. The shaded band marks the 12-hour switch window; the bold white line is always actual ISO-NE demand at the logged observation time.
⏳ Loading forecast time-series data…
Time Window
—
Data Points in View
—
Miners Shown
—
Best Forecaster (by latest MA score)
—
2025
Accomplishments
- 60+ students engaged
- 1 deployed decentralized AI subnet
- Production workflows live on GCP
- Onboarding time reduced by >90%
- Cross-institution collaboration
- UConn is the first institutions to offer experiential decentralized AI learning
- Partnership with Yuma established for mentorship and operational support
- Ready infrastructure for developing custom ML models
- End-to-end governance system implemented
2026
Vision
Formalizing Innovation
Offer hackathons and final projects for UConn graduate students.
Industry Alignment
Deepen the Yuma partnership and explore a strategic collaboration with SCORE on decentralized computer vision infrastructure built on Bittensor.
Experiential Learning
Hackathon Day: live deployment sessions to solve real-time infrastructure challenges.
System Architecture
High-Level System Overview
Validator-Miner Ecosystem Architecture
The complete operational architecture features three concurrent validator loops:
Validator
Miner 1
Miner 2
Miner N
0
Minutes Interval
0
Seconds
Function: Queries miners every 5 minutes with timestamp-based challenges
prediction_scheduler() → neurons/validator.py
0
Seconds Delay
0
Minutes
Function: Processes predictions after a 10-minute delay, allowing actual load values to materialize.
evaluation_loop() → neurons/validator.py
0
Seconds Interval
0
Minutes
Function: Updates network state by syncing with the blockchain
metagraph_resync_scheduler() → neurons/validator.py
Miners run Axon servers on port 8091 that handle Challenge synapses by fetching ISO-NE five-minute system load data (LoadMw).
Core Prediction Workflow
Prediction-to-Reward Lifecycle
The two-phase lifecycle with delayed evaluation:
1
Phase 1: Collection
Collects predictions every 5 minutes and queues them with timestamps
Every 300s
2
Phase 2: Evaluation
Processes predictions after a 10-minute delay, allowing actual load values to materialize before evaluation.
After 10 minutes
3
Reward Calculation
Calculates scores
Prediction=12985.725571428571, Actual LoadMw=13025.843, Reward=0.0214
Meet the team
Faculty, leads, and core contributors building BittBridge AI at UConn.
- Dr. David Wanik Faculty Advisor
- Dmitrii Tuzov Product & Tech Lead
- Faeze Safari Lead ML & Data Systems Architect
- Benjamin Bosco System Engineer Lead
- Matthew Bernstein Data science & Research Lead
- Niharika Sharma Project Manager Lead