A selection of engagements we've shipped — across fintech, healthcare, e-commerce, SaaS, and government. Names anonymized where required by NDA.
The team was burning $190k/month on GPT-4 for 22M monthly support requests. We replaced with a distilled Llama 3.1 70B fine-tune on vLLM, added semantic + KV cache, and routed cheap queries through Haiku 3.5. Eval pass rate held within 1.2pp.
Private VPC on AWS Bedrock + OpenSearch vector. Full audit trail, model cards, LLM-as-judge nightly eval against 1,400 gold answers. Passed two third-party security reviews and an internal red-team exercise.
Replaced batch scoring with online inference using Triton + ONNX on G6 GPUs. Feature store on Tecton with online/offline parity. CTR up double-digits, GPU spend flat through autoscaling.
LangGraph supervisor pattern. Specialist agents for billing, account, integrations. Tool registry via MCP. Confidence-gated escalation with full trace replay. CSAT moved from 4.2 to 4.6 / 5.
Air-gapped Llama 3.1 70B on bare-metal H100 cluster. STIG-hardened K8s, Cosign-signed artifacts, PII redaction at ingress and egress. ATO achieved on first audit.
End-to-end documentation: model cards, system cards, FRIA (Fundamental Rights Impact Assessment), continuous monitoring. Decision-explainability surfaced to claimants and adjusters.
Migrated Spark ETL to Iceberg + Trino + dbt. Forecasting models on Ray, served via SageMaker async. Pipeline cost down 71%, training cycle from 14h to 2.3h.
Hardened the entire ML lifecycle: signed model artifacts, dataset lineage, immutable training records, role-scoped serving credentials. Walked auditors through every control with engineering present.
Compliance, latency, accuracy, and cost trade-offs differ by domain. We've worked across enough of them to bring patterns, not surprises.
Tell us what you're building. If we can help, we'll come back with a written approach and an estimate within five business days.
Start the conversation →