Yatharth L.

I take on AI and data problems that are hard because of scale or ambiguity: agents that call tools and act on real systems, streaming data pipelines, and observability for AI products that work in a demo but fail in production. A decade of backend engineering, founding engineer at three startups (one from zero to Series C), currently Technical Lead at Spocto X (Yubi). Bring me in when: - Your voice bot, chatbot or agent is live, but you can't see why it fails or measure the metric you sell on - You need an agent that plans, calls tools and writes to real systems, with guardrails, evals and cost control - Your event volume has outgrown your database and you need a streaming pipeline with real-time analytics - Non-technical teams need to query operational data in plain English - Your LLM bill or latency keeps climbing and nobody can explain why - You want an independent audit of an AI system's retrieval, prompt-injection exposure and cost Selected work: Voice AI observability (large voice-AI company, 20,000-30,000 calls/month) The team couldn't measure their containment rate. I built the data pipeline and an observability dashboard covering containment rate, human transfers, call volume as a funnel, LLM call volume, LLM and third-party failures, intent-level flow breakdowns, and A/B test results. On top of it, a natural-language search that combines text-to-SQL, embedding search and deterministic rules, so the team can ask questions without writing SQL. Within the first month they could see where calls were failing, and containment rose from 7% to 9%. Real-time streaming pipeline (Yubi, collections data) Kafka ingesting 250,000 requests per minute into ClickHouse, giving stakeholders a real-time view of system performance. Cut cost by about 81% with P95 read latency under 200ms. Real-time analytics platform (Spocto X), built solo end to end About 100M records, 1-10M new records/day, dashboards across ~50 user-selected filters at P95 under 100ms. DuckDB streams Parquet straight into ClickHouse, so data never passes through the service. Financial AI agent (founding engineer, Amalgamic) RAG with tool calling and n8n orchestration. The MVP helped the startup raise $100K+ and runs for about $100/month. Health-tech, zero to Series C (Ayu Health, first founding engineer) - Calling system: replaced a Redis scheduler that had become a single point of failure with an event-driven architecture handling 200,000+ events/day at P95 under 500ms; 30,000+ calls/day, pickup rate up 25-40% - Document AI: 30,000 documents/day, processing time from 5 hours to 1 hour, LLM spend from $5K to $2K/month - MySQL: read replicas for analytics plus query and index work, about 10x database performance - Cloud cost: AWS down 40%, GCP Vision/OCR down 80% Agent tooling A Claude Code plugin that generates, validates and repairs DevOps configs; published MCP servers; CLAUDE.md conventions and guardrails for engineering teams. How I work: I scope before I quote and tell you when a simpler tool is the right answer. Builds are priced by milestone. Ongoing data and AI-quality work runs as a monthly retainer. Not a fit: If you're still validating an idea, no-code and AI builder tools will get you an MVP faster and cheaper than I can. I come in once it's proven and needs to become a real system. Fastest way to start: a 60-minute paid consult where we work on one real problem. If it's a fit, we scope the engagement from there. Stack: AI: Claude, OpenAI, LangChain, AI agents, tool calling, MCP, RAG, evals, text-to-SQL, embeddings Data: Kafka, ClickHouse, DuckDB, PostgreSQL, MySQL, Redis, Elasticsearch, Celery Backend: Python, FastAPI, Node.js, Java, Go Voice and automation: Retell, Twilio, n8n, Make Cloud: AWS, GCP, Docker, CI/CD, New Relic