Ternary ALU & TensorFlow Integration
Budget: £20 – £250 GBP
I am building an experimental futures-trading pipeline that hinges on a custom 3-trit (base-3) ALU and a TensorFlow model trained directly on its state transitions. The work is about creating each piece in isolation and more about making them talk to each other seamlessly.
What I already know I need
• A fully specified ternary ALU that supports arithmetic, logical, and shift operations.
• A clear interface (ideally HDL-level plus a Python bridge) so the ALU’s inputs, internal states, and outputs can be streamed in real time.
A way to train ths ALU in python, passing backend triton kernels in ternary fp32
• A TensorFlow training script that ingests those streams, builds the first proof-of-concept predictive model, and saves checkpoints for later tuning.
Acceptance snapshot
1. The ALU passes a testbench covering every arithmetic, logical, and shift path.
2. A Python driver pushes at least 10 000 random trit vectors through the ALU and captures results without mismatch.
3. The TensorFlow notebook trains on that data and outputs a loss curve plus a saved model file.
If you have prior experience with non-binary logic or have already toyed with ternary HDL cores, you will feel right at home here. Feel free to suggest the most comfortable HDL (Verilog, VHDL, Chisel, or even a software emulation first); I only care that the integration with TensorFlow is clean and reproducible. Source code, brief documentation, and test assets are the only deliverables I’m expecting.
What I already know I need
• A fully specified ternary ALU that supports arithmetic, logical, and shift operations.
• A clear interface (ideally HDL-level plus a Python bridge) so the ALU’s inputs, internal states, and outputs can be streamed in real time.
A way to train ths ALU in python, passing backend triton kernels in ternary fp32
• A TensorFlow training script that ingests those streams, builds the first proof-of-concept predictive model, and saves checkpoints for later tuning.
Acceptance snapshot
1. The ALU passes a testbench covering every arithmetic, logical, and shift path.
2. A Python driver pushes at least 10 000 random trit vectors through the ALU and captures results without mismatch.
3. The TensorFlow notebook trains on that data and outputs a loss curve plus a saved model file.
If you have prior experience with non-binary logic or have already toyed with ternary HDL cores, you will feel right at home here. Feel free to suggest the most comfortable HDL (Verilog, VHDL, Chisel, or even a software emulation first); I only care that the integration with TensorFlow is clean and reproducible. Source code, brief documentation, and test assets are the only deliverables I’m expecting.