Turning complex business operations into intelligent systems

Transcription, status detection, and LLM evaluation of every call against 10+ operator KPIs — script adherence, objection handling, empathy, upsell technique, resolution quality. Disputed cases go to a human arbitration loop whose decisions feed back as training signal, which lifted accuracy a further 15%. Migrating from a third-party LLM API to a self-hosted 70B model cut running costs roughly in half at the same throughput.

Fixed dialling settings are a bad trade: dial fast and callers wait on hold, dial slowly and operators sit idle. This controller solves the allocation on a rolling horizon instead, modelling transfer delay as an empirical lag convolution and validating through discrete-event simulation calibrated to production data. Results shown are from counterfactual replay against the previous policy. The method is published.

Prospecting runs without a human in it: the system discovers businesses through the Places API, classifies them by category and fit, calls to qualify, then ranks by warmth from engagement signals and sends a personalised follow-up. Full call history is retained, so a prospect who calls back is resumed in context rather than re-introduced cold. A parallel pipeline discovers and negotiates influencer partnerships, selecting the commission structure per audience profile automatically.

A single static risk score can't tell a coordinator what to do this morning. This system runs five horizon-specific models over point-in-time features — using only what was knowable at each decision point — then surfaces the highest-risk training sequences and ranks candidate students for recovery by availability, prior disruption, training need, and current risk. Validated chronologically, with historical replay and what-if analysis built into the coordinator workflow.

Documents are classified and routed by a vision-language model with OCR fallback, then answered through a hierarchical RAG pipeline with sentence-level citation and recursive three-level summarisation. The entire stack was migrated off third-party APIs to self-hosted deployment to satisfy the client's data-residency and confidentiality requirements.

There is no single source of truth for addresses across 15+ countries. Each country has its own authoritative providers alongside Google Places, and each region and city carries its own formatting conventions and validity rules — so the system holds a per-country source set and a regional rule set rather than one global model.

Disagreements between sources are surfaced with plain-language reasoning instead of a bare confidence score, so a reviewer can spot-check in seconds. The most useful part is the loop: the LLM acting as judge proposes changes to the rules themselves when it sees a pattern the current rules mishandle, and a human confirms or rejects them. The rule set improves per region without anyone rewriting it by hand.