All studios are deployed as permanent, zero-cold-start static web applications on Hugging Face Spaces (sdk: static) with interactive client-side execution:

# 1. Profile local server capacity (Ollama, vLLM, llama.cpp) with automated roofline tuning
pip install servebench && servebench quick

# 2. Pre-flight analytical capacity planning, multi-GPU matrix & cloud $/1M tokens (zero GPU needed)
pip install infersize && infersize compare --model qwen2.5-7b --quant fp16

# 3. Continuous-batching inference runtime with Paged KV-cache (pure PyTorch, zero C++ build steps)
pip install mini-infer

# 4. Discrete-event serving simulator with Radix prefix caching & chunked prefill
pip install sched-sim && sched-sim run --scheduler chunked --chunk-budget 512

# 5. From-scratch quantization engine & Pareto evaluation benchmark (INT4/INT8/FP8/BitNet 1.58b)
pip install quantbench-engine && quantbench pack-inspect --format int4

# 6. Multilingual token inflation tax analyzer across 52 languages
pip install token-atlas && token-atlas compare --text "Hello world"

Flagship	Core Architecture	Empirical Silicon Numbers	Release & Provenance
mini-infer	Lightweight continuous-batching LLM runtime in pure PyTorch for Qwen2.5. Iteration-level dynamic scheduler, Paged KV-cache with dynamic sub-byte INT4/FP8 compression (-72.7% VRAM), radix prefix caching, and Leviathan speculative verifier.	• 27.0 tok/s at c=4 (+130.8%) • 2.04 Tok/Joule (13.2 W) • -72.7% KV-cache VRAM (INT4) • 68/68 tests pass (100%)
servebench	LLM serving capacity load-tester & auto-tuning harness (servebench tune). Multi-trial repeats with 95% CIs and synchronized NVML GPU telemetry.	• Honest p99 ($N \ge 100$) • MBU roofline profiling • Slot/Queue/Bandwidth diagnostics • 169/169 tests pass (100%)
quantbench	Unified from-scratch quantization engine (INT4/INT8/FP8/BitNet 1.58b), vectorized sub-byte packing, Ampere 2:4 structured sparsity, dynamic QuantLinear layers, and dual-metric Pareto evaluation benchmark.	• 3.5x INT4 / 7.5x BitNet • 0-copy hook dequant • 53/53 tests pass (95% cov)
sched-sim	Discrete-event LLM serving simulator with first-principles roofline latency models, SGLang-style Radix prefix caching, Sarathi chunked prefill, and Perfetto trace visualizer.	• <4.8% silicon error • 0.87s full sweep • 45/45 tests pass (87% cov)
infersize	Pre-flight analytical performance modeling, multi-GPU comparison matrix, speculative decoding sizer, and cloud TCO ($/1M tokens) capacity planner. Interactive Web Studio, NVML hardware fit verifier, and dynamic Hugging Face Hub ingestion.	• <10 ms CPU solve • 60/60 tests pass (89% cov) • Zero GPU required
token-atlas	Multilingual token inflation tax visualizer and cost calculator across 52 world languages and 6 foundation tokenizers.	• 52 languages analyzed • 24/24 tests pass (100%) • 0ms static runtime
runsonmine	Community database of measured (not estimated) LLM inference speeds on real consumer silicon (laptops, desktops, Macs).	• 24 empirical runs • 24/24 tests pass (100%) • 100% SLA attainment
EdgeSLM	Air-gapped document sensitivity classifier and Tier-1 DLP secret scanner for edge CPU inference.	• 20+ HF downloads (v3) • 6.7× dynamic batching • 51/51 tests pass (100%)

I believe that engineering credibility is built by documenting trade-offs, architectural failures, and hardware bounds alongside performance gains:

• Naive KV-Allocation Fails Under Dynamic Lengths:
  Pre-allocating contiguous tensor buffers for worst-case context lengths wasted >60% of GPU memory due to internal and external fragmentation. Fixed by implementing fixed 16-token physical block pools with virtual-to-physical block tables in pure PyTorch, cutting memory waste to <3.8%.
• Unbatched Iteration Bottlenecks on DRAM Memory Bandwidth:
  Stepping concurrent sequences through serial model calls forced the GPU to stream model weights from DRAM 4 separate times per step (capping decode at 11.7 tok/s on an RTX 3050). Switching to zero-copy physical block gathering streams weights exactly once per step across all sequences, jumping throughput to 27.0 tok/s (+130.8%) and delivering 2.04 Tok/Joule.
• Head-of-Line Prefill Starvation Under Continuous Batching:
  In standard FCFS continuous batching, a 2,048-token prompt prefill monopolizes the compute pipeline for >120ms, causing active decoding streams to stutter with severe TPOT latency spikes. Validated through sched-sim that Sarathi-style Chunked Prefill ($C=512$) interleaves prompt chunks with decodes, keeping TPOT under 40ms without hurting throughput.
• Transformer Layer Quantization Sensitivity Asymmetry:
  Attention Query ($W_q$) and Key ($W_k$) projections suffer 1.8–2.4 dB greater SQNR degradation under low-bit quantization than feed-forward MLP layers because phase drift in the attention dot-product $QK^T / \sqrt{d_k}$ propagates exponentially through softmax. Demonstrated in quantbench that keeping $W_q, W_k$, and $W_{\text{down}}$ in higher precision or finer group size ($G=32$) yields near-FP16 perplexity while compressing MLP Gate and Up into aggressive sub-byte formats.
• No Cherry-Picked Latency Metrics:
  Single-run p99 latency spikes are statistically deceptive without sufficient sample size. servebench strictly enforces $N \ge 100$ before calculating p99 percentiles, and reports 95% confidence intervals across multi-trial repeats.

• Roofline Benchmarking & Auto-Tuning LLM Serving: How We Pushed Pure PyTorch to 27 tok/s (2026-10-23)
  Measuring empirical Memory Bandwidth Utilization (MBU), saturation knees, and multi-stream batching to eliminate TTFT degradation on consumer laptop GPUs.
• Why Naive KV-Caches Waste >60% VRAM (And How Paged Memory Fixes It) (2026-10-22)
  How replacing contiguous memory allocations with 16-token physical blocks and virtual table mapping brings production PagedAttention to pure PyTorch.

Runtime Systems   : Paged KV-Cache Allocators · Continuous Batching · Radix Prefix Caching · Speculative Decoding · Discrete-Event Simulation
Silicon Profiling : Roofline MBU Modeling · NVML Telemetry · Saturation Knee Detection · TTFT / ITL / TPOT Analysis
Core Stack        : PyTorch (SDPA, Tensor Gathers) · Python 3.10+ · CUDA / NVML · ONNX Runtime INT4 · Hugging Face
Quantization      : Symmetric AbsMax · INT4 Blockwise · BitNet 1.58b Ternary · Ampere 2:4 Structured Sparsity · SmoothQuant · AWQ / GPTQ
Serving & Infra   : FastAPI · Asyncio / Aiohttp · SSE Streaming · vLLM · Ollama · Docker · GitHub Actions CI/CD
Security & Trust  : PyPI Trusted Publishers (OIDC) · Sigstore Cryptographic Attestation · Zero-Secret Airgap DLP

• Portfolio Website: prajwal-pujari.vercel.app
• Hugging Face Hub: huggingface.co/prajwalpujari16
• PyPI Packages: pypi.org/user/Prajwal2002PI
• LinkedIn: linkedin.com/in/prajwal-pujari-mn9964
• Email: prajwalpujari16@gmail.com