I BUILD SYSTEMSTHAT DON'T BREAKUNDER load.
FIG.01Harsh Kamoriya — software engineer building backend systems, distributed infrastructure and AI-powered products.
§01 — ENGINEERING SIGNAL
Knight badge. Contest-tested, not tutorial-tested.
Graphs, DP, segment trees — the boring foundations.
17 hours of compute collapsed to 56 minutes.
End-to-end response time halved in production.
Provisioned, deployed, monitored, debugged.
I like problems that get harder when you add users.
I work across backend systems, distributed infrastructure and AI applications — from architecture and implementation to deployment, debugging and production operations. The part I enjoy most is the part after launch: when real traffic finds the bottleneck you didn't plan for.
Production, under pressure.
- PROBLEM↓
- ARCHITECTURE↓
- BOTTLENECK↓
- OPTIMIZATION↓
- RESULT
A trading platform choking on its own success.
Mintzy's analytics pipeline computed signals per ticker synchronously. As the watchlist grew, users waited a full minute for a response.
Every layer, touched.HOVER A LAYER TO INSPECT
VERVIQ.
AI interviews that actually listen.
The challenge: LLM interviewers forget. Ask a follow-up three minutes in and the context is gone. The solution: a multi-agent loop where a dedicated memory agent and RAG keep every answer in play.
Speaks over WebRTC. Audio + transcript stream in real time.
Gemini-driven interviewer runs the conversation and scores answers.
Distils every answer into structured memory — claims, gaps, signals.
Pinecone retrieval over the résumé, JD and prior answers.
A context-aware question that only a human interviewer would think to ask.
gitsaathi
Point it at a repository. It reads the code, the commit history and even meeting recordings, and tells you where the real complexity lives.
VIEW LIVE ↗AEROGUIDE
An AI airport companion that works when the terminal Wi-Fi doesn't — on-device Whisper transcription with graceful fallbacks.
VIEW LIVE ↗
→ Start from the failure modes, not the happy path.
→ Load-test it until it tells you where it's weak.
→ Logs, traces, lock graphs. Find the cause, not the symptom.
→ Measure first. Then make the slow part fast.
→ Into production, with CI/CD and someone watching the graphs.