B5 Polish: Enterprise Features
- Added '--dry-run' flag for cost estimation (FinOps).
- Implemented RICE Scoring logic within the prompt.
- Refactored the CLI into an Express.js API for scalability.
Working log
Dated entries, written while it was happening
Nothing here is edited after the fact. Where an entry turned out to be wrong, the correction appears as a later entry rather than a rewrite of the earlier one.
B5 Build: The PM Bot
- Implemented the CLI tool using Node.js and Commander.
- Built the 'Reflexion Loop' for automated quality control.
- Integrated Zod to force INVEST-compliant JSON output.
B5 Learn: Agentic Patterns
- Studied the ReAct pattern (Reasoning + Acting).
- Learned strict JSON Schema definitions for Function Calling.
- Designed the 'Triple Threat' architecture (Cluster -> Context -> Story).
B4 Polish: The API Wrapper
- Built a Node.js API to wrap the Python inference model.
- Created a 'Golden Set' of tickets to benchmark performance.
- Calculated final ROI: Buy (GPT-4) vs Build (Llama-3).
B4 Build: Training Llama-3
- Ran the training job on Google Colab T4 GPU.
- Optimized hyperparameters (Learning Rate, Epochs) to prevent overfitting.
- Hit hardware limits and pivoted to Hosted Inference Endpoints.
B4 Learn: Fine-Tuning Theory
- Studied LoRA (Low-Rank Adaptation) and QLoRA.
- Learned the difference between 'Pre-training' and 'Instruction Tuning'.
- Prepared the dataset: Formatted CSVs into Alpaca JSONL format.
B3 Polish: Trust & Citations
- Solved hallucination issues by forcing the model to cite sources.
- Added a UI component to expand and read the source document.
- Deployed the app to Streamlit Cloud.
B3 Build: The Chat Interface
- Built a frontend using Streamlit for rapid prototyping.
- Connected the LangChain backend to the Vector Store.
- Implemented 'Conversation Memory' so the bot remembers context.
B3 Learn: RAG Architectures
- Studied the Retrieval Augmented Generation (RAG) pattern.
- Learned LangChain fundamentals: Chains, Retrievers, and Loaders.
- Mapped out the architecture for a Corporate Handbook Chatbot.
B2 Polish: Tuning Retrieval
- Experimented with different Chunk Sizes (256 vs 512 tokens).
- Added metadata filters (e.g., filter by 'Date' or 'Author').
- Visualized vector clusters using PCA.
B2 Build: The Search Engine
- Built an ingestion script to chunk long documents.
- Generated embeddings using
text-embedding-3-small. - Wrote a SQL Remote Procedure Call (RPC) to match vectors in Supabase.
B2 Learn: Vector Math
- Deep dive into Embeddings: How text becomes numbers.
- Learned Cosine Similarity vs. Euclidean Distance.
- Set up a Supabase Vector database.
B1 Polish: Prompt Optimization
- Refined prompts to enforce specific formatting (Bullet points vs Paragraphs).
- Calculated token costs for batch processing.
- Documented the tool for non-technical team members.
B1 Build: Scripting Logic
- Wrote my first Python script to loop through a CSV of meeting notes.
- Connected script to GPT-3.5-turbo for cost-effective summarization.
- Handled basic API errors (Rate Limits) with try/catch blocks.
B1 Learn: The API Paradigm
- Transitioned from ChatGPT UI to OpenAI API usage.
- Learned core parameters: Temperature, Top-P, and Max Tokens.
- Studied 'System Prompting' frameworks to control AI persona.