Design notes from Echo Echo's lyrics pipeline: a CrewAI agent whose prompt budget scales with track duration, importing a heavy dependency optionally, normalising section tags for a downstream model, and chaining generation backends so one failing service doesn't break the app.
A lyric generator is a good example of a task where "the model can do it" is the easy part and constraining the output so the rest of the system can use it is the work. This post covers how Echo Echo's lyrics step is set up, and how it fits with the music generators it feeds.
The lyrics step is a single CrewAI Agent (role: Lyricist, allow_delegation=False) given a single Task, run through a one-agent Crew. It would be easy to over-build this into a multi-agentA system in which several LLM-driven agents, each with its own role or tools, collaborate on a task and pass work between each other. graph; a single agent with a precise task is simpler, cheaper and easier to debug. The model is a small, fast Groq-hosted Llama model, chosen by checking whether a Groq key is present.
The task prompt is built from track metadata — title, mood, genre, BPM, duration, theme, instruments and the music prompt — followed by hard rules:
The line budget is computed from the duration, not left to the model: 8 lines for tracks up to 30 seconds, 16 up to 60, 24 up to 90 and 48 beyond. The same number appears in both the instructions and the expected_output. Numbers computed by code and repeated in the prompt are far more reliable than adjectives like "short". An empty response raises a specific error instead of silently returning blank lyrics.
CrewAI is a large dependency. The module imports it inside a try/except ImportError and sets the names to None if missing; each function that needs it raises a typed LyricsGenerationError("CrewAI is not installed."). A lean cloud deployment can then ship a smaller requirements file (the repo has separate deploy and full requirements) and the app still starts — the feature just reports that it's unavailable. The same pattern is used for the optional Gradio client and for local MusicGen.
LLMLarge language model — a neural network trained on large amounts of text to predict the next token, used for generation, summarisation, classification and more. return section headers in many forms: [Verse 1], [CHORUS], Pre-Chorus:, Hook. The ACE-Step vocal model expects a specific lowercase tag format. _format_lyrics handles it in layers: regexes collapse numbered or mixed-case headers to [verse], [pre-chorus], [chorus], [bridge] and so on (mapping "hook" to chorus); if the text is already in the right shape it's returned unchanged; otherwise a line-by-line pass converts plain-text section headings; and empty input becomes a safe instrumental placeholder. A cheap, deterministic adapter between two probabilistic components is one of the best reliability investments.
The app has more than one way to produce audio — a hosted music API, a hosted open model reached through a Hugging Face Space, and an optional heavy local model — and a commit near the end fixed the local model fallback and the lyric refinement path. The shape is the useful part: each backend is a separate module with the same kind of inputs, timeouts are configured per backend (the Space client uses a long HTTP timeout and the duration is clamped to a valid 10–240 second range), and failures are caught at the boundary so the user gets a message instead of a stack trace.
Generating a BRD With Parallel Agents — Frozen Snapshots, Explicit Gaps and a Last-Step Summary
The structure of Beacon's document generator: a frozen snapshot of classified signals, six section agents running in a thread pool with progress events, an executive summary that runs last, and a rule that missing data produces an explicit placeholder instead of invention.
Separating Signal From Noise in Project Chatter — A Two-Phase Heuristic and LLM Classifier
How Beacon's noise filter works: cheap regex and rule heuristics decide the obvious cases in parallel, only ambiguous chunks go to batched Groq LLM calls, a confidence threshold routes uncertain labels to human review, and a local keyword fallback keeps the pipeline alive without an API key.
Streaming an LLM Reply Through Django — SSE, Session Memory and the Save That Didn't Happen
Notes from building a small Django chatbot: proxying a streaming model API with StreamingHttpResponse, why session memory must be saved manually after a stream, how multi-chat state fits in the session, and what changed when free model limits kept shifting.