Luis Lacerda

Legal lead generation and LegalTech software for motor-vehicle accident, SSDI, VA disability and mass-tort practices.

• Direct a standing team of AI coding agents (Claude Code, Grok CLI, Gemini CLI, Cursor cloud agents). I write the specifications, test what they produce, and approve each change through pull requests.
• MTOS CRM (mtoscrm.com), TypeScript on Node.js: a lead-distribution CRM for mass-tort firms. Each buyer has its own validator; delivery is sent automatically only when a lead meets that buyer's criteria, and duplicates and criteria-not-met leads are tagged rather than lost. It integrates with buyers' Litify instances, captures TrustedForm consent certificates, publishes an OpenAPI specification, and is held to more than 1,000 automated API tests.
• MTOS Portal (mtosportal.com): a vendor portal with tracked referral links, per-vendor attribution, and lead status that updates live from the CRM.
• Case Closed FL: a Next.js personal-injury platform. Built the lead validators for SSDI, Case Closed FL and F4MR traffic.
• tcpa-firewall: a TCPA list-scrubbing service that removes non-compliant records before any outreach begins.
• Launched about twelve legal lead-generation sites, among them caseclosedmva.com (Arizona motor-vehicle accident intake) and caseclosedfl.com. Each uses a multi-step qualifying form, is built for search and AI answer engines, scores about 99 on Lighthouse, and runs on DigitalOcean App Platform.
• Manage the lead-vendor network and LinkedIn outreach to law firms, and apply TCPA, CAN-SPAM and CASL requirements to both.

Online advertising and SaaS.

• Apply business, data and financial analysis to advertising and software engagements.

An ongoing supervised fine-tuning run on a 36,000-step schedule, carried out on one RTX 4070 Ti with 12 GB of VRAM.

• Base model: Qwen3-14B (abliterated variant). Base weights are frozen and quantised to 4-bit NF4; low-rank adapters train with bf16 compute using Hugging Face transformers and PEFT.
• At step 10,127, roughly 28% of the schedule, I merged the adapter into the base weights mid-run, converted the checkpoint to GGUF and served it through Ollama, so interim evaluation could proceed while optimisation continues.
• Training data: a ~1.9 GB JSONL instruction/response corpus with a held-out validation split.
• The corpus includes frontier-coding-sft-v1, a 2,640-pair component I curated in-house across eight domains, including advanced coding, software engineering and cybersecurity.
• Retrieval over the same corpus: bge-small-en-v1.5 embeddings indexed in Chroma, top-6 retrieval, exposed through an OpenAI-compatible proxy that supports tool use.
• Supporting work: directed a PyTorch training guide that runs from MNIST to LoRA, QLoRA and GGUF conversion; set up Flux LoRA training in ComfyUI; operate local inference with Ollama, LM Studio, ComfyUI/Flux and Kokoro TTS.

• Designed and operate a fleet of more than ten coordinated AI agents covering coding, code audit, recruiting, LinkedIn, Microsoft Teams monitoring, Google indexing and media. Scheduled routines assign the work; watchdogs catch what stalls.
• Grok-Memory: a Git-based shared memory and retrieval repository that every agent reads from and writes to, with a companion retrieval project, GENERAL_RAG.
• BOS-Omega, an architecture for orchestrating several LLMs, and AURA-Runtime, an agent runtime written in Rust.

• LegalHeatmapApp (C#/.NET 8).
• Arabic speech-recognition backend (quran-fastconformer).
• Robotics arm work (SO-ARM100, LeRobot).