Hey, I am Harsha — an engineer who builds systems and measures them: how models are served efficiently, how agents are made reliable enough to trust, how both behave under adversarial pressure, and what the infrastructure underneath — gateways, ledgers, matching engines — actually costs when you benchmark it honestly.

Most of my recent work is in projects — 75 shipped systems across five engineering tracks, every one a working implementation with committed benchmark reports:

• Machine learning & AI systems (15) — LLM serving internals, agent reliability, multi-agent protocols, AI security, retrieval and memory, cost, and post-training. Highlights: PageServe (continuous batching and a paged KV cache, token-for-token exact), Durable Agents (kill -9 mid-task, resumes exactly-once), and InjectGuard (prompt-injection containment by construction, 100% → 0% attack success).
• Forward-deployed engineering (15) — making agents function inside real organizations: Enterprise Agent Deployment (approval gates defined over effects, not tool names), Legacy MCP (0/6 → 6/6 task success through a governed wrapper over an authentically awful SOAP backend), and DocPipe (a document pipeline that verifies its own extractions).
• Software engineering (15) — developer infrastructure for the agent era: Sandboxd (the same hostile suite run on gVisor and Firecracker, so isolation and latency are compared rather than asserted), RepoCtx (repo-scale context that lifts an 8B agent’s file recall 3.4x), and MiniTemporal (a durable-execution engine in ~2k lines of stdlib Go, chaos-verified).
• HFT & low latency (15) — latency engineering and market microstructure: Tick2Trade (583 ns p99 internal with a published per-hop budget), Matchbook (a zero-unsafe Rust matching engine within 1.01-1.09x of pointer-based C++, verified by 2.49B fuzz executions), and Deadman (a kill switch model-checked in TLA+, with the Go core replaying all 23,240 model sequences).
• Backend engineering (15) — the service layer AI products stand on: LLM Gateway (1.84 ms p99 added latency, cost ledger exact to the picodollar), Meterd (292,792 chaos deliveries reconciling exactly), and Flowd (durable execution surviving 165 worker SIGKILLs with zero lost or duplicated effects).

A habit that runs through all of them: build the measurement harness before drawing the conclusion, report the number the harness produced — and keep the honest negatives.

Interests

1. LLM inference and serving efficiency
2. Agent reliability, evaluation, and durable execution
3. Low-latency systems and market microstructure
4. Backend infrastructure for AI products — gateways, metering, multi-tenant isolation
5. Information retrieval and recommendation systems
6. The intersection of ML and privacy

Publication

Data Efficient Subset Training with Differential Privacy — arXiv:2503.06732

I am also passionate about open source and contributing to the community. You can find my code on GitHub.