Open AI engineering developer meetups, architecture jams & demo sprints:

What was built:

Engineered and deployed collaborative multi-agent runtimes with deterministic tool routing, persistent memory state machines, and human-in-the-loop review gates. Attendees built functioning autonomous agent networks executing live code, validating structured schemas, and recovering gracefully from tool execution failures.

Key takeaways:

• State machine architectures in LangGraph: managing cyclical agent loops without token bloat
• Deterministic structured output extraction: enforcing strict Pydantic JSON schemas with zero runtime crashes
• Multi-agent supervisor routing: delegating sub-tasks across specialized worker agents with contextual memory

What was built:

Built and deployed an end-to-end local streaming inference service pairing an 8B target vision-language model with a 1B speculative draft verifier. Implemented KV-cache reuse, benchmarked token throughput (142 tok/s achieved on consumer RTX hardware), and exported production GGUF checkpoints for local execution.

Key takeaways:

• KV cache reuse mechanics: eliminating redundant prefill passes across speculative draft steps
• Draft acceptance rate dynamics: balancing draft model temperature to maximize multi-token speculation verification
• Quantized vision encoder latency trade-offs: why FP8 weights outperform 4-bit AWQ on cross-attention projections

What was built:

Built resilient autonomous workflows using Gemini structured outputs and tool definitions. Engineered deterministic execution loops that query external APIs, parse typed schema responses, and enforce recursion depth guardrails.

Key takeaways:

• Constraining LLM outputs with Pydantic JSON schemas to guarantee 100% parse rate in production pipelines
• Designing idempotent tools and recursion depth tokens to prevent catastrophic infinite billing loops
• Adding human-in-the-loop gates and rollback semantics for high-stakes operational side-effects

What was built:

Built an end-to-end production retrieval engine over complex technical PDFs. Implemented semantic boundary chunking, dual indexing (BM25 sparse + dense embeddings), reciprocal rank fusion (RRF), and cross-encoder re-ranking to crush semantic hallucinations.

Key takeaways:

• Why naive cosine vector similarity fails on exact technical keyword queries and multi-column balance sheets
• Cross-encoder latency trade-offs: when to re-rank top 30 vs top 100 for optimal sub-200ms p95 response time
• Synthesizing test queries with Ragas to quantify precision recall, context relevancy, and faithfulness

What was built:

Every developer cloned the community repo, quantized a Llama-3-8B base model into 4-bit NF4, applied rank-16 LoRA adapters, fine-tuned on domain-specific markdown in under 35 minutes, and exported merged GGUF weights for local execution.

Key takeaways:

• VRAM allocation math: why 16GB is plenty for 8B models with gradient checkpointing & flash attention
• Loss curve inspection: catching catastrophic forgetting and learning rate divergence in the first 50 steps
• Adapter merging & GGUF serialization: zero-overhead local deployment on Ollama and llama.cpp