Bob: Offline HVAC AI (Powered by StarFlame)

Field-hardened HVAC AI (Gemini 1.5 Pro). StarFlame compresses memory 24% for offline use. Features 1.5x Safety & Jitter Reconstruction.

Comment

Inspiration

I am a co-owner of an HVAC company, and I work in the field every day. The biggest problem with modern AI is that it assumes you have a fast internet connection. In reality, technicians work in basements, industrial freezers, and rooftops where the signal is zero.

Cloud-based AI is useless to me when the connection drops. Local LLMs are often too small to handle complex diagnostics safely. I needed an intelligence that was dense, offline, and safety-prioritized. I didn't just want a chatbot; I needed a second brain that could fit on my phone.

What it does

Bob is a field-hardened HVAC assistant running on our custom StarFlame Architecture (built on Gemini 1.5 Pro).

Instead of storing massive, linear conversation logs (which bloat memory and require the cloud), Bob uses Metric-Guided Lossy Summarization. He compresses complex troubleshooting interactions into mathematical "State Summaries"—tiny, vector-based isotopes that retain the logic of the solution while discarding the noise of the conversation.

This allows Bob to:

  1. Run Efficiently: We achieved a 24% reduction in memory footprint compared to standard logging.
  2. Prioritize Safety: Bob identifies dangerous scenarios (like high-pressure lockouts) before suggesting repairs.
  3. Operate in Low-Signal Zones: The compressed knowledge base is small enough to be cached locally on standard Android hardware.

How we built it

We built the core logic using Google AI Studio and Gemini 1.5 Pro. The "StarFlame" architecture is a custom prompting and data-handling kernel that enforces specific constraints on the model's context window.

The Math Behind StarFlame

We treated memory as an engineering problem, not a storage problem.

  • Compression: We utilized Semantic Folding to calculate the vector centroid between the Problem ($A$) and the Fix ($C$), discarding the conversational fluff ($B$). $$Entropy Rejection Ratio \approx 24\%$$
  • Reconstruction: To prevent the "hallucinations" common in high-compression systems, we implemented a Gaussian Noise Injection during retrieval. $$Jitter(\sigma) = 0.30$$
  • Safety Weighting: We hard-coded a multiplier to all "Safety" and "Sentiment" vectors before compression to ensure emotional context isn't lost. $$Safety Weight = 1.5x$$

Challenges we ran into

The biggest challenge was the trade-off between Compression and Empathy. When we first compressed the logs, the AI became "cold" and purely logical—it forgot the customer was frustrated.

  • Initial Emotion Fidelity: 8.54% (Catastrophic failure)
  • Solution: We implemented the 1.5x Emotional Weighting protocol. This forces the model to treat "User Frustration" as a hard technical metric, preserving it alongside the error codes.

Accomplishments that we're proud of

  • 24% Compression: Proving that we can fit a Senior Technician's knowledge base into a fraction of the space.
  • 99% Logic Fidelity: The system retains technical accuracy even after aggressive summarization.
  • The "Jitter" Discovery: Finding that adding noise ($\sigma = 0.30$) actually improved the stability of the memory retrieval.

What's next for Bob

We are currently integrating Bob directly into the Method Media Worker App (our field management software). The next step is Read-Only Archival, creating immutable "State Summaries" to prevent long-term data drift. Bob is getting smarter, denser, and safer with every job.

Built With

Share this project:

Updates