📍 Dubai, UAE
I build AI systems for places where a wrong answer costs money.
Agentic platforms, MCP servers, and eval infrastructure for enterprise banking. 7+ years shipping production AI: ISRO → Swiggy → Mensa Brands → banking-scale platforms. ~7k stars on open source. Writing read by 80k+ engineers.
AI System Design Guide · ★ 1.7k How to design production AI systems and evals. Actively maintained. This is where my current thinking lives.
FAANG Coding Interview Questions · ★ 5.1k The curated interview prep list, forked 1.2k+ times.
TranscriptAI Multilingual speech-to-text CLI on OpenAI Whisper. Batch processing, 8+ formats.
Screwdriver CI/CD 10+ merged PRs across UI, models, and core. Shipped GitHub/GitLab PR validations used by 10k+ engineers.
| System | The number | Context |
|---|---|---|
| Fintech Document AI | $0.01 per doc | 200k requests/min at 99.5% accuracy, 90% cost reduction |
| Trade-docs compliance agents | weeks → hours | review time down from 1-3 weeks to 2-4 hours, 98% audit completeness |
| Agentic Engineering Platform | 33% infra saved | MCP orchestration with hierarchical tool taxonomy, 57% compute optimization |
| DiffusionID portrait pipeline | 95% cheaper | than commercial APIs, at 450 images/hour and 10k+ cards per batch |
| Core platform reliability | 99.9% uptime | up from 88%, queries cut from 45s to 200ms, $2M+ downtime saved |
Writing: The Real Signal
No-hype takes on AI, written from inside production systems:
- Uber Blew Its AI Budget in One Quarter. So Will You.
- Nobody Goes to Parties and Talks About TCP/IP (on MCP quietly winning)
- Karpathy Says He's Never Felt This Behind. That Should Scare You.
- You Don't Pick the Best Model. You Pick the One That Gets You.
AI/ML: Claude, GPT-4, Llama 3.1 · SDXL, InstantID · LangChain, PyTorch, Hugging Face · RAG, agentic workflows, vector search Backend: Python, Java, Go, TypeScript · PostgreSQL, Redis, DynamoDB, Elasticsearch · Kafka, SQS, EventBridge Infra: AWS (Bedrock, SageMaker, EKS, CDK) · Kubernetes, Terraform · Datadog, Prometheus, NewRelic
def ship_ai_system(requirements):
"""
Rules I learned in production, the expensive way:
1. Build the meter before you turn on the tap
2. No evals, no deploy
3. Treat LLMs as unreliable services; design for failure
4. The boring fallback saves you at 3 a.m.
5. Agents need governance more than they need autonomy
"""
if handles_money_or_compliance(requirements):
return validation_layers() + audit_trails() + human_in_the_loop()
if needs_scale(requirements):
return async_pipelines() + caching() + cost_meters()
return something_simple_that_works()Researching multi-agent orchestration and eval infrastructure. Mentoring engineers on distributed systems and AI system design. Consulting with teams taking agents from demo to production.