Research
Experiments, field notes, and open research questions.
Devanagari Subword Tokenization Expansion
"Can customized Devanagari BPE rules reduce sequence length by >40% without increasing perplexity?"
Result: Subword merge table reduces Nepali token count from 24 to 11 on sample benchmark.
Async KV Cache Pruning in Low-VRAM GPUs
"How much KV cache can be safely evicted without losing contextual coherence in long dialogs?"
Result: Attention-head importance scoring allows 30% eviction with minimal quality loss.
Probabilistic Type Checking in AI-Native AST Compilers
"Can compiler type systems verify output schema compliance at compile time for Blyx?"
Result: Proof-of-concept type checker asserting JSON output structure bounds.
Distributed P2P GPU Scheduling
"Can latency-aware scheduling achieve 80% cluster utilization over consumer networks?"
Result: 4-node topology achieving 18–120ms ping routing.
Language → Robot Motion Control (HTN Planning)
"Can language prompts decompose into deterministic HTN motion goal predicates?"
Result: 5-stage pipeline mapping instructions to motor control signals.
Why KV cache matters in multilingual inference
Key-value caching is one of those infrastructure details that most people overlook until they hit a wall with inference latency. For multilingual models serving low-resource languages, caching strategies need to account for varying token densities across scripts.
What makes an AI-native language different?
Most programming languages are designed for deterministic execution. An AI-native language needs to handle uncertainty, context, and generation as first-class concepts. The compiler primitives are fundamentally different.
Why distributed GPU compute is hard
The hard part isn't acquiring GPUs. It's the networking, peer scheduling, fault tolerance, and making it all work reliably across heterogeneous hardware in markets where connectivity is unpredictable.
HTN planning vs end-to-end VLA in robotics
End-to-end VLA models are elegant, but hierarchical task networks with language-constrained sub-agents provide better safety bounds and real-time debuggability.