Fintech / RAG · compliance2.4M documents
RAG copilot for compliance analysts
Task. Analysts spent 30–60 minutes on each case review, searching 2.4M internal documents and checking regulatory requirements. Open-source search returned irrelevant results, and naive RAG hallucinated.
What we did. Hybrid search (BM25 + embeddings) with cross-encoder re-ranking. Chunking that follows document structure. A separate fact-checking pass on Claude. A citation-aware UI where every claim links to its source. Evaluation on a golden set of 1200 pairs.
Result. Faithfulness 0.94 (against a 0.71 baseline), case review time down from 40 to 6 minutes, token cost 72% lower thanks to prefix caching and routing simple questions to cheaper models.
6.5×faster analysis
0.94faithfulness
−72%cost vs naive RAG
E-commerce / agents800K tickets / month
AI agent in a marketplace support team
Task. A high volume of standard requests: "where is my order", "refund me", "change the address". First-response time was 9 minutes, support staff were burning out, NPS stood at −18.
What we did. An agent pipeline on LangGraph with tool use: order status check, return initiation, address update, escalation. Human-in-the-loop for refunds above a threshold. A separate safety layer against prompt injection.
Result. First response went from 9 minutes to 12 seconds, 68% of requests are resolved automatically, support NPS is +12. The support team moved to complex cases and proactive outreach.
45×faster first response
68%auto-resolution
+12 NPSsupport
SaaS / LLMOps10M req/day · multi-team
Production LLM gateway with routing and caching
Task. 14 product teams were calling OpenAI and Anthropic independently. Cost grew faster than revenue, nobody could see who was spending what, rate-limit failures were frequent and there was no fallback.
What we did. Designed and launched a gateway: a unified API, routing between Claude, GPT and Llama by task type and SLO, a semantic cache, per-tenant rate limiting, retry and fallback on 429 / 5xx, cost attribution via OpenTelemetry. PII filters and a per-user token budget.
Result. Token cost down 64% in a quarter. Gateway overhead of 42 ms at p99. Zero rate-limit incidents in 6 months. Every team sees its spend in real time.
10Mrequests / day
−64%token cost
42 msoverhead p99
DevOps / AI agent1.4K incidents / month
AI agent for automatic triage of production incidents
Task. The SRE team was drowning in alerts: 50% false positives, MTTA of 18 minutes, on-call engineers woken 3–4 times a night.
What we did. An AI agent reads the alert, pulls the related metrics and logs over MCP, searches past incidents in postmortems (RAG), and proposes a hypothesis and a runbook. Below a confidence threshold it escalates to a human with the context already gathered.
Result. MTTA went from 18 minutes to 2 minutes, because the agent gathers the context for the engineer. 41% of alerts are closed without on-call as known issues. Night-time wake-ups are down 3.5×.
9×lower MTTA
41%auto-triage
−72%on-call burden
Analytics / data lake & DWH120K call hours / month
LLM-powered analysis of sales calls
Task. Call recordings were unstructured noise. Listening to them manually was impossible, the legacy BI covered only the CRM, and managers filled in fields inconsistently.
What we did. A pipeline: transcription (Whisper), diarization, structured extraction (an LLM writing to a strict JSON schema), highlights of objections and key questions, ClickHouse views, dashboards for the sales team and for product.
Result. 100% of calls are tagged automatically within 5 minutes of the call ending. The product team sees real customer objections, and managers get feedback with quotes. Sales conversion grew 9% in a quarter.
100%call coverage
5 minSLA from end of call
+9%conversion