Lead Engine: Social-Only Lead Sourcing with Compiled Local Models
An autonomous lead pipeline where the model observes and Python calculates, with the scoring rubric compiled into local Ollama models.
GPU-aware two-tier models that yield to video renders, knowledge files compiled into four model variants, a deterministic five-gate F.I.R.E. rubric owned by Python, seed-pinned scoring for reproducibility, and review-then-send drafts.
- Job boards removed after 377/385 leads proved uncontactable
- Rubric compiled into models; Python owns all arithmetic
- Seed-pinned scorer gives byte-identical runs
- GPU-aware model tiering yields to render jobs
Freelance lead generation usually means scraping job boards. This engine found that 377 of 385 job-board leads were unusable because "apply on the site" is not a contact route, so it was rebuilt around social sources: Reddit multireddits, public Telegram channel previews, and dormant YouTube and X connectors. The people worth talking to are the ones stuck on a problem ("my backtest doesn't match live"), not the ones posting job ads.
The model layer is pragmatic about a shared GPU: a 24B model runs when the GPU is free and a 7B CPU-only model takes over when a video render owns it, and overnight render batches are never interrupted. "Training" means compiling: knowledge files are baked into four Ollama model variants (lead and outreach for each size) at a cost of a few megabytes thanks to content-addressed weights.
Scoring is the design centre. An LLM asked for a holistic score clusters high and drifts, so the model only observes evidence against five gates and four banded dimensions, and Python owns every number. A fixed seed makes runs byte-identical so calibration is real. Outreach drafts are reviewed before sending and the writer may never claim an engagement absent from the profile.
How it works
- Sources are polled for new posts matching the ICP
- Model extracts evidence quotes against the five gates
- Python computes dimension bands, modifiers and tier
- High-tier leads get a drafted outreach message
- Owner reviews and sends; outcomes feed calibration