> narun@console:~$

Demos don't close deals. Production does.

Voice AI · LLM · PCI & SOC — 4 years shipping at YC startups.

Let's talk about your problem

You're here because:

Your MVP needs to ship in 6 weeks. Your team is 8 weeks away.
You raised a round. Now you need to actually build the thing.
Your "AI feature" is a demo. It needs to ship.
Compliance is blocking your enterprise deal. PCI. SOC2.
Your infra is duct tape. It works until it doesn't.
You patched the security hole. But the architecture is still exposed.

I fix these.

$0K+

monthly debt recovery

0K+

monthly calls & chats

0%

latency reduction

arXiv

2410.17950

What I build

MVP Sprint

01

6-8 weeks. Idea to deployed product. Tech stack that scales. No technical debt.

IDEAL FOR: Pre-seed with funding, no technical co-founder

Book MVP Call

AI Integration

02

Voice AI. LLM pipelines. ASR systems. Function calling. Production-grade, not demo-grade.

IDEAL FOR: Seed+ startups adding AI features that need to actually work

Book AI Call

Compliance & Security

03

PCI DSS. SOC2 prep. Security hardening. The stuff blocking your enterprise deals.

IDEAL FOR: Startups blocked on enterprise sales by compliance

Book Compliance Call

Architecture Review

04

Deep dive into your codebase. What breaks at scale. How to fix it. Kubernetes. Infra. State machines.

IDEAL FOR: Series A+ needing external validation before major decisions

Book Review Call

Where I've shipped

Rifa

Founding Engineer

Rebuilt the voice bot from scratch on state machines and WebSockets — 50% lower latency, 100K+ monthly calls. Designed PCI DSS card capture with a BERT classifier (no generative model in the data path). Built in-house ASR processing 25K calls/day.

  • Rebuilt voice bot on state machines and WebSockets — 50% lower latency, 100K+ monthly calls
  • PCI DSS card capture with BERT classifier — no generative model in the data path
  • In-house ASR processing 25K calls/day
  • Graph-based testing framework: HDBSCAN-driven test case generation from production conversations, structured LLM tool call logging, static/dynamic flow validation
  • Live credential compromise → restructured attack surface with VPN-gated access controls, hardened perimeter; no further security incidents
  • GCP migration, Kubernetes deployment with full observability (OpenTelemetry, SignOz), reducing infrastructure downtime
  • Evaluated 10+ TTS models → Kyutai streaming model eliminated buffering entirely: 800ms → 250ms latency
$500K+/mo recovery100K+ calls/mo70% latency cut25K calls/day

Floworks (YC W23)

Founding Engineer

Built the function calling execution engine (arXiv 2410.17950) — parallel/async tool orchestration benchmarked against OpenAI and Anthropic. Earlier, built the WhatsApp platform that attracted $250K in LOIs and contributed to YC W23 acceptance.

  • Function calling execution engine (arXiv 2410.17950) — parallel/async tool orchestration benchmarked against OpenAI and Anthropic
  • WhatsApp platform that attracted $250K in LOIs and contributed to YC W23 acceptance
  • Autonomous cold email and LinkedIn agent: campaign execution, multi-database lead enrichment, follow-up sequencing, >2% meeting booking rate at scale
  • Led the technical pivot from WhatsApp to cold outreach — built the campaign, sequencing, and meeting-booking infrastructure
  • AWS infrastructure at 99.95% uptime under SOC compliance, Terraform IaC, GitHub Actions CI/CD
arXiv 2410.17950YC W23$250K LOI pipeline99.95% uptime

Limited availability

Let's build something together

Currently available for projects
Schedule a 30-min call

No pitch. No BS.
Just: your problem, my take, whether I can help.