I lead teams by day and build ecosystems by night.

My stack runs from Android NDK to bare-metal Rust services — no abstraction layer is too high, no system too low. I've fine-tuned LLMs, written custom WebRTC SFU servers, shipped OS-level Android tooling, architected microservice migrations for Fortune 500 energy infrastructure, designed post-quantum cryptography platforms, and somewhere in between, started questioning whether transformers are the right architecture at all.

Primary languages: Kotlin · Rust. Secondary religion: making things fast.

Not side projects. A connected vision: own every layer of the AI development lifecycle — and secure it while you're at it.

AI writes fast. Runlit keeps score. An eval layer that catches what code review misses — hallucinated APIs, intent mismatches, security vulnerabilities, and compliance violations — before they reach production.

$ runlit check --pr 2847
→ scanning 147 lines (AI attribution: Cursor/claude-sonnet)
  hallucination     0 phantom APIs found          ✓  1.00
  intent            matches spec, minor drift      ✓  0.91
  security          no critical issues             ✓  0.97
  compliance        1 medium finding (PCI 6.2.4)   ⚠  0.88
  eval.score        94 / 100
✓ CLEARED — merge when ready.

139 rules. 12 provider-specific packs (OpenAI, Anthropic, LangChain, Stripe...). 30-second GitHub App install. Merge-blocking. MIT-licensed rules.

1.7× more issues in AI-generated code. 45% of AI tasks introduce security vulnerabilities. 75% of enterprise engineers will use AI coding tools by 2028. Runlit exists because those three numbers are a collision course.

Rust GitHub App GitLab Azure DevOps Bitbucket CLI GitHub Actions

Intelligent LLM routing layer between your app and 10+ GPU providers. Scores every provider every 30s on price, latency, uptime, and queue depth. Routes to the cheapest option that meets your SLA — <8ms overhead, <200ms auto-failover.

# Before
client = openai.OpenAI(api_key="sk-...")

# After. That's it.
client = openai.OpenAI(api_key="rtr-...", base_url="https://api.routra.dev/v1")

Typical savings: 60% of inference spend. Adaptive ML routing via LinUCB bandit — learns per model, region, and time-of-day.

Rust Axum OpenAI-compatible Python SDK TypeScript SDK Go SDK Rust SDK CLI

Decorate a function. Get a production HTTPS endpoint backed by GPU hardware. No Docker, no YAML, no DevOps hire. Not Python-first — full feature parity across Python, TypeScript, Go, and Rust. Same mental model, idiomatic syntax per language.

Serverless GPU Python TypeScript Go Rust Multi-Language SDK

Not a chatbot. A background-capable multi-agent operating layer — voice-first, connector-driven, deeply personalised via persistent per-user memory (imprint.bit). Orchestrates Gmail, Calendar, Notion, Slack, GitHub, and custom MCP servers. Every external dependency behind a swappable adapter trait.

SENSA (intent) → ORACLE (task decompose + DAG) → Agent Mesh → Connectors → ECHO (response)

Rust Axum Tauri 2 Next.js 16 Claude (Anthropic) pgvector Whisper ElevenLabs

Route storage, compute, and data workloads across AWS, GCP, and Azure — automatically. Built for Indian businesses: INR billing, UPI/Razorpay, GST invoices, DPDP compliance scanning. Phase 1 shipped: gRPC-first architecture, AES-256-GCM per-tenant encryption, Rust adapters, full SDK surface.

Rust gRPC Protobuf Next.js PostgreSQL 18 Redis Razorpay

The full-stack post-quantum cryptography migration platform. Secures all four layers — transport, authentication, secrets storage, and credentials — under one open core, before quantum computers make today's encryption obsolete.

1Password secures passwords. Cloudflare secures transport. Nobody secures all four layers under one brand with one open core. QShield is that platform.

Built in Rust. Aligned to NIST FIPS 203 (ML-KEM), FIPS 204 (ML-DSA), and FIPS 205 (SLH-DSA). India-first. Enterprise and government-ready. Harvest-now-decrypt-later attacks are already happening — the migration window is open and won't stay that way.

Rust ML-KEM ML-DSA SLH-DSA PyO3 NAPI-RS WASM Tauri 2 Flutter Apache 2.0

Neural Execution & Understanding System. Not an LLM variant. A new architecture class.

NEXUS addresses fundamental flaws in the transformer architecture. NOOR is the first model family built on it — scaling bottom-up with intention, not brute force:

Also running Grafts — NEXUS modules injected into existing open-source backbones (Qwen-NOOR, Gemma-NOOR) as experimental sidelines.

This is what happens when you spend enough time staring at transformer limitations that you decide to do something about it.

Systems & Backend

AI & Infra

Mobile & Multiplatform

• Application Lead @ Vendeep / NRG Energy (2022 – present) — Leading a 25-person team modernizing legacy enterprise portlets into microservices. 145+ APIs migrated across Spring Boot, FastAPI, and Salvo. Still running.
• Senior Android + Reverse Engineering @ Innobuzz (2020 – 2022) — NDK-level engineering, custom OS injection, Android security tooling at the system layer.
• Delivered 4 production phases for HCTRA (Harris County Toll Road Authority, Houston) — Kafka streaming, Oracle GoldenGate, millions of daily toll transactions.
• Built and fine-tuned a generative AI training platform for HCTRA — Meta Llama 3, Rust + Python, custom fine-tuning pipeline.