Praveen
I build practical, inspectable tools around open models: adapters, quantization workflows, evaluation surfaces, and small demos that make model behavior easier to understand.
Featured work
PathPack-Q · LFM2.5-2.6B training-free quantization
An architecture-specific post-training quantization experiment for LiquidAI's hybrid convolution/attention LLM. PathPack-Q uses exact gated-path channel permutations to improve which weights share each 4-bit quantization group, without training, text calibration data, extra parameters, or runtime operators.
- 5.30% lower perplexity than byte-matched uniform MLX 4-bit on 8,160 held-out WikiText-2 tokens
- 9.87% lower KL divergence from the BF16 teacher on a fixed prompt suite
- Identical 1,517,616,892-byte checkpoint and 4.501 effective bits/weight
- Includes the search algorithm, complete-path acceptance gate, rejected-layer evidence, checkpoint builder, and machine-readable evaluations
ScopeGuard · Qwen2.5-1.5B LoRA
A locally trained agent decision-layer adapter that turns natural-language requests into strict JSON risk and confirmation decisions before tools execute.
- 93% risk accuracy on a 100-example held-out split, up from 66% for the base model
- 100% exact schema compliance, up from 84%
- 3.957M trainable parameters — only 0.256% of the 1.5B base model
- Includes adapter weights, original dataset, deterministic generator, training config, baseline outputs, and per-example evaluation
→ Explore the ScopeGuard dataset
→ Open the complete benchmark explorer
LoRA Lens
An in-browser audit tool for adapter_config.json files. It surfaces rank, alpha, scaling, target modules, reproducibility gaps, and conservative parameter-efficiency estimates without uploading weights or requiring an API key.
Current lab
- LoRA and PEFT adapter design
- Quantization and memory-aware inference
- Reproducible model cards and evaluation tooling
- Human-readable demos for technical work
I prefer falsifiable, transparent experiments with clear limits over opaque claims.