Systems we’ve built, measured in outcomes

Challenge

Reliance on commercial LLM APIs for reasoning and tool-calling agents created mounting cost, latency, and data-control exposure as usage scaled across multiple business lines.

Approach

Designed and delivered an end-to-end agent factory: synthetic conversational data generation, LLM-as-judge quality scoring, fine-tuning, evaluation, and serving — productized as a config-driven, multi-tenant platform rather than a one-off model.

Outcomes

• Platform adopted as the default path for new model-training initiatives across 6+ business lines
• Fine-tuned a 32B-parameter reasoning model that matched or exceeded baseline tool-calling accuracy at materially lower inference cost than commercial APIs
• New-pipeline setup time reduced from 3–4 weeks to under a day

Challenge

Fragmented, surface-specific detection models created inconsistent precision and blind spots across a payments ecosystem spanning core payments, peer-to-peer transfer, gateway, buy-now-pay-later, identity, and checkout.

Approach

Architected a unified anomaly-detection framework with shared feature and evaluation standards, deployed independently per surface but governed by one quality bar, layered with automated decision-intelligence tooling for analyst workflows.

Outcomes

• Greater than 90% precision sustained across all six transaction surfaces
• Multi-agent analytics for natural-language querying and automated validation reduced manual analyst time by more than 30%

Challenge

High-quality labeled data was the primary bottleneck for fine-tuning production models — manual annotation cycles measured in weeks could not keep pace with model iteration.

Approach

Built a multi-agent annotation platform combining independent LLM policy-agent labelers, a judge model for consensus, and classical ML annotators across nine annotation types, with statistical agreement scoring and confidence-based routing between auto-accept, spot-check, and deep review.

Outcomes

• Large-scale labeling cycles cut from weeks to hours
• 70–80% of manual review eliminated through automated confidence routing
• Labeled corpora fed directly into downstream fine-tuning pipelines

Challenge

Training data for a production reasoning agent needed to cover more than one hundred real-world intents with realistic variation, at a volume manual curation could not support.

Approach

Designed a chain-of-thought synthetic data pipeline built on a large-scale teacher model, with a multi-dimensional persona model for distributional realism, two-stage embedding-based deduplication, and human-in-the-loop validation gated by an LLM-as-jury benchmark.

Outcomes

• 87,000+ examples generated per run across 109 production intents in under 45 minutes end-to-end
• 5–10% semantic duplicates removed automatically before reaching training
• Established the training-data foundation for a flagship reasoning agent now in production

Challenge

Executive and operational teams lacked real-time, trustworthy visibility into KPIs, OKRs, and product health — most reporting was manual, backward-looking, and inconsistent across teams.

Approach

Delivered a multi-agent decision-intelligence stack for natural-language-to-SQL reasoning, automated validation, and summarization; layered a retrieval-augmented executive assistant over structured and unstructured data; shipped rule-bound recommendation engines for compensation and performance workflows.

Outcomes

• Manual data-task time reduced by more than 30%
• Compensation and performance recommendation engine achieved 100% rule compliance across five career levels
• Real-time OKR and product-health visibility delivered to executive teams