realtime voice + gtm infrastructure. turning calls into a continuously learning outbound system.
one thing i'd immediately explore at retell is turning voice interactions into a continuously learning gtm system instead of treating calls like isolated events.
most outbound systems today are still primitive: static sequences, fixed scripts, generic lead scoring, disconnected crm updates.
but voice agents generate an insane amount of high-signal conversational data.
i'd build infrastructure that captures and operationalizes that data in realtime.
retell agents run thousands of outbound calls. instead of just logging transcripts, the system should cluster objections dynamically, identify conversion-driving phrases, rerank lead quality continuously, generate new outbound angles automatically, recommend follow-up timing, and create high-performing script variants from successful calls.
basically: the outbound system itself becomes self-improving. not just ai making calls. ai improving how calls happen.
retell feels interesting because it sits at the intersection of realtime systems, ai infra, human behavior, distribution, and operational scale.
which is basically the kind of environment i naturally enjoy operating in.
i'm not someone who wants to maintain static funnels or run repetitive growth playbooks.
i like building systems that:
python, sql, typescript, flutter
without really separating them.
example:
someone visits pricing page 3 times. retell agent calls within minutes: references company context, understands use-case, qualifies budget/timeline, books meeting automatically, pushes structured intel into crm.
not chatbot-level automation. actual operational automation.
a lot of this thinking also came from building trxnd.
trxnd started from a simple belief: most products don't fail at the build. they fail at distribution.
so i built systems around:
because gtm to me feels less like marketing and more like systems engineering.
most projects sit somewhere between ai infra, realtime systems, automation, operational leverage, and gtm infrastructure.
multi-agent Android automation system. Controls phones the way a human would: reads the accessibility tree, decides what to tap next, no APIs needed. Automates job applications, handles CAPTCHAs by handing off to the user, accepts commands via voice/PDF/Telegram, supports 15+ languages. Built on OpenClaw + DroidRun + FastAPI + React.
real-time voice-cloning video translator. Takes a voice sample, then during a live call transcribes, translates (Gemini), and speaks back in your cloned voice via Qwen3-TTS. Sub-545ms end-to-end. WebRTC + Redis Pub/Sub.
LLM inference engine built from scratch. Dynamic batching (20ms window), KV-cache rewrite from O(N²) to O(1) per token, HTTP + gRPC, Prometheus/Grafana observability, distributed round-robin router.
agentic RAG pipeline for causal extraction from conversational data. LangGraph orchestration, adaptive reranking, LLM judges. F1 = 0.94. Placed 5th at Inter-IIT Tech Meet 14.0.
AI mental wellness platform. Real-time conversational avatar via Gemini audio APIs, emotion-based music generation via Lyria, RAG-powered Persona Dashboard. Scaled to 1,000+ concurrent users.
at saturn labs, hundreds of hours of robotics training data were being manually qc'd every day.
people were checking:
manually.
so i built and integrated an automated qc pipeline overnight. the next day the team stopped wasting engineering time on repetitive review work.
that's usually how i operate.
i think the most exciting thing about retell is that voice is still massively under-optimized.
most companies still treat calls like support overhead, sales overhead, operational overhead.
when really they're one of the richest behavioral data layers inside a business.
and whoever operationalizes that layer best probably wins. that problem space feels very fun to me.