Hack nights, workshops & demo days:

What was built:

Every developer cloned the community repo, quantized a Llama-3-8B model into 4-bit NF4, applied LoRA adapters, and fine-tuned on custom datasets using Unsloth in under 35 minutes.

Key takeaways:

• VRAM allocation math: why 16GB is plenty for 8B models with gradient checkpointing
• Loss curve inspection: catching catastrophic forgetting in the first 50 steps
• Adapter merging: exporting GGUF weights for instant local Ollama execution

What was built:

Built an end-to-end retrieval system over complex technical docs. Implemented chunking with semantic boundaries, dual indexing (BM25 sparse + dense embeddings), and cross-encoder re-ranking.

Key takeaways:

• Why naive vector similarity fails on exact technical keyword queries
• Cross-encoder latency trade-offs: when to re-rank top 30 vs top 100
• Synthesizing test queries with Ragas to quantify precision recall

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 log trace telemetry.

Key takeaways:

• Constraining LLM outputs with Pydantic JSON schemas to guarantee 100% parse rate
• Designing idempotent tools to prevent catastrophic duplicate mutations
• Adding human-in-the-loop gates for high-stakes operational side-effects