Quiet Signals
Lab

A functional sales or support agent needs two things: a clear model of the conversation it should have, and a reliable way to interpret what the user actually means at each step. I build both — the business logic as a deterministic state machine, with language model components inserted only where probabilistic behaviour is genuinely wanted (intent matching, paraphrase, extraction).

The architectural payoff is operational integrity. The agent cannot say something it wasn't authorised to say, every transition is logged, and the system's correctness doesn't move when the underlying model is updated by a vendor. This matters most where conversations carry liability — regulated industries, customer data flows, anything subject to audit.

Quiet Signals Lab operates a verified Meta messaging application that client businesses connect their Facebook Pages and Instagram accounts to. Rather than building a new integration per client, businesses get access to a single tested and approved system.

Deliverables:

• Custom agent architecture with configurable business rules
• Integration with messaging platforms (Meta, WhatsApp, Telegram)
• State management and fallback handling
• Human escalation routing
• Documentation and deployment support

The part of a RAG system most likely to fail is retrieval. A model can only work with what it's given; if the retrieved context is wrong or poorly ranked, the answer is wrong regardless of which model generates it. I build retrieval pipelines where chunking, embedding selection, and reranking are treated as the core engineering problems, and where every generated response carries provenance back to the source passages it was grounded in.

For workflows where data can't leave the client's infrastructure, the entire stack runs locally — Ollama or vLLM for inference, ChromaDB or pgvector for storage, no third-party calls in the answer path. For less restrictive contexts, commercial providers (OpenAI, Mistral) substitute in cleanly. The architecture stays the same; only the inference target changes.

Deliverables:

• Production-ready RAG backend with REST API
• Document ingestion pipeline and vector database setup
• Retrieval optimisation: chunking, embedding tuning, reranking
• Source attribution and generation constraints
• Documentation and handover

Many regulated workflows still rest on people reading documents and copying fields into systems. The bottleneck is rarely volume — it's that the documents don't follow predictable schemas, and the cost of a wrong field (a misread date of birth, an inverted clause, a missed citation) is high enough that pure automation fails the reliability bar without further structure.

I build pipelines that combine deterministic structure — OCR, layout parsing, defined-term and cross-reference resolution — with schema-constrained LLM extraction and per-field confidence scoring. Outputs route automatically: high confidence to acceptance, mid to human review, low to rejection. Every extracted field carries a provenance link back to its source span, and every pipeline event is logged for audit.

The deterministic layers are built on Contractex, the open-source library released through Quiet Signals Lab.

Deliverables:

• End-to-end pipeline from raw document to structured output
• Per-field confidence scoring with routed human-in-the-loop
• Provenance tracking with character-level source spans
• Audit-ready event logs for GDPR-compliant processing
• Integration with downstream CRM, filing, or PDF workflows

Production models fail in two registers. The familiar one: distribution shift, data quality issues, patterns that didn't generalise. The less familiar one, more relevant in regulated settings: memorisation of training data, susceptibility to extraction attacks, behaviour under adversarial input. I run controlled ablation studies on the first set and integrity probes on the second, so fixes can be prioritised by evidence and the system's failure surface is documented before it ships rather than after a security review surfaces it.

Deliverables:

• Root cause diagnosis with quantified impact analysis
• Distribution shift analysis and ablation studies
• Memorisation and extraction surface testing
• Documented failure modes and monitoring thresholds
• Recommendations ranked by priority