tap
TRACK 01

Hack Nights: Live Code in the Room

Every session starts with a verified GitHub repository. We clone, configure dependencies, understand the math and memory constraints, and run inference or fine-tuning runs before leaving the room.

Hardware: 1x RTX 3090/4090 or M3 MaxIntermediate

Local LoRA Fine-Tuning

Stack: Unsloth, PyTorch, Hugging Face PEFT

Quantize base 8B models to 4-bit NF4, apply rank-16 LoRA adapters, fine-tune on custom technical markdown in 35 minutes, and export merged GGUF weights.

Throughput: 850+ tokens/secAdvanced

Production vLLM Inference Engine

Stack: vLLM, PagedAttention, Ray, Docker

Deploy continuous batching, chunked prefill, and OpenAI-compatible streaming endpoints on Linux instances with automatic GPU memory paging.

Recall: 94.2% • p95: 180msIntermediate

Hybrid RAG: BM25 + Dense Vectors

Stack: Qdrant, BM25, BGE-reranker-large, FastEmbed

Construct a two-stage retrieval engine combining lexical keyword search with dense embeddings and reciprocal rank fusion to eliminate semantic hallucination.

Validation: 99.8% Valid JSONBeginner–Intermediate

Autonomous Agent Tool-Use Runtime

Stack: Gemini 2.0 Flash API, Pydantic, Instructor

Build deterministic multi-turn tool loops with typed schema validation, automatic retry on truncated brackets, and human-in-the-loop operational gates.

Zero Cloud Cost • Sub-50ms TTFTAdvanced

Local Edge AI with ONNX & WebGPU

Stack: Transformers.js, WebGPU, ONNX Runtime

Run quantized small transformers directly inside client browser tabs and desktop apps with zero backend server roundtrips.

Self-Instruct DecontaminationIntermediate

Synthetic Dataset Bootstrapping

Stack: Python, Datasets, LlamaIndex, vLLM

Generate high-quality domain-specific instruction datasets using frontier models with automated deduplication and semantic contamination checks.

TRACK 02

War Stories: What Failed at 2 AM

Unfiltered retrospectives on production crashes. We inspect real trace telemetry from silent vector drift, runaway API billing loops, and p99 inference spikes, breaking down exactly how to engineer resilient guardrails.

Incident #01 • Agent OperationsRoot Cause Analysis

The $14,000 Infinite Recursive Tool Loop

Stack: Autonomous Agents, Tool Calling

Detailed post-mortem on how an unchecked agent spawned 4,200 nested sub-queries in 38 minutes, and the recursion depth tokens that permanently fixed it.

Incident #02 • Retrieval QualityArchitecture Review

Chunking Size Disasters in Financial Docs

Stack: RAG, Markdown Parsers, Chunking

Why fixed 512-token windows sliced multi-column balance sheets in half and generated 40% error rates in financial summaries.

Incident #03 • InfrastructureInfrastructure Tuning

Cold-Start Inference Spikes on Serverless

Stack: AWS Lambda, Modal, RunPod, PyTorch

Diagnosing 28-second cold starts with 7B models on serverless containers and migrating to persistent pre-warmed GPU pools.

Incident #04 • Search PipelinesData Pipeline

Vector Drift After Embedding Updates

Stack: Vector DB, Embeddings, Evaluation

An accidental model version bump that silently degraded retrieval recall to 41% without triggering a single HTTP 500 error.

Incident #05 • Production ReliabilityReliability Engineering

JSON Output Truncation on Edge Cases

Stack: Structured Decoding, Outlines

Hard token limit crashes that severed closing braces, and why grammar-based constrained decoding saved production pipelines.

Incident #06 • PerformancePerformance Tuning

Memory Leaks in Async Model Handlers

Stack: PyTorch, CUDA Cache, FastAPI

CUDA memory fragmentation in Python async workers that forced daily container restarts and the precise torch.cuda.empty_cache() fix.

TRACK 03

Paper Sprints: Research into Working Code

We read preprints the week they drop and build minimal, clean Python and CUDA proof-of-concepts to test whether novel architectures deliver their promised throughput and latency gains.

Speedup: 2.8x Higher ThroughputResearch to Code

Speculative Decoding with Draft Models

Stack: PyTorch, Draft Verification Kernels

Pair a compact 1B draft model with an 8B target model to verify multiple candidate tokens per forward pass and slash latency.

Complexity: O(N) Linear TimeCutting Edge

Mamba & State Space Architectures

Stack: Mamba-SSM, PyTorch CUDA Kernels

Analyze selective state spaces, compare memory footings against quadratic attention, and benchmark throughput across 64K context tokens.

Reasoning Scaling LoopsDeep Dive

Test-Time Compute & Search Trees

Stack: Monte Carlo Tree Search, Verifiers

Build multi-path candidate generation with reward verification models to solve complex algorithmic tasks during inference.

Unified Encoder AlignmentExperimental

Vision-Audio Cross-Modal Projection

Stack: Qwen2-VL, Whisper, Cross-Attention

Pass raw spectrograms and visual patches directly to unified decoder transformer heads without converting to intermediate text.

Zero Reward Model OverheadAdvanced

Direct Preference Optimization (DPO)

Stack: TRL, Hugging Face, PyTorch

Align models with pairwise preference loss directly on chosen vs rejected answers without running unstable PPO reinforcement loops.

Top-2 Expert Dynamic GatingArchitecture

Sparse Mixture of Experts (MoE) Routing

Stack: MoE Routing Layers, DeepSeek/Mixtral

Implement auxiliary load-balancing loss, route tokens through top-k feedforward experts, and measure compute efficiency.