selected work
Not side projects — platforms with real users, real traffic and real numbers. Built at Stashfin, Zupee and Dresma AI.
A self-serve platform for building and deploying production AI agents — prompt, tools, model and channels configured from a UI. Cut agent delivery from days to minutes. Today it runs multiple production agents, including a customer-support agent handling ~30,000 messages and 44M tokens per day across 6,000+ daily users.
agent delivery: days → minutes · 44M tokens/day
One gateway for every LLM call — budget tracking, rate limiting, multi-provider fallbacks and circuit breaker patterns so product teams survive provider outages without noticing them.
300K+ req/day · 15–20ms · 99% uptime
LangGraph multi-agent architecture for response planning and conflict resolution, RAG memory retrieval with Qdrant, and dynamic response timing that simulates human conversation patterns.
46% D15 retention
Turns any REST API into an agent-callable tool: configure the request, fire a live test call to capture the real response shape, annotate the keys — and the LLM tool schema writes itself. Hand-written integrations: eliminated.
any API → agent tool, zero code
End-to-end knowledge ingestion — upload, chunk, embed, store — plus a contact system unifying each user's email, phone and Slack identities into a single record with persistent memory across channels.
custom knowledge bases without engineering
Parallel processing of 50–100 microseries for ad creative, plus a no-code bot management console giving product managers 2-click control over models and behavior.
1 week → 3 hours production time
Near-fail-proof architectures for heavy computation workloads on AWS with Kafka-driven microservices — engineered for stability under load.
−40% response time · +50% stability
A practical write-up on self-hosting open-weight models — hardware, quantization and serving trade-offs from hands-on experimentation.