Systems plumbing, concurrent pipelines, and secure runtime isolation for agent startups. The claim I'd rather be judged on: not that my agents are correct, but that being wrong is discoverable. Write-up: https://sriramvarun0636.github.io/
Vasool — compliance-gated payment recovery agent on Razorpay's test APIs. I pre-registered seven falsification criteria before running anything, and then lost against one: a dumb baseline that retries everything recovers 16.35pp more than mine. It also breaks policy in 1,000 of 1,000 seeded runs; mine breaks it in 0. That trade is the finding, and it's the second section of the README, not an appendix. The LLM never calls a tool — inert verdict type, no adapter to the execution plane, asserted by an import-graph test. Simulated outcomes, tagged as such in the source; every figure is a key in a committed manifest, so you can check any number without running anything.
https://github.com/sriramvarun0636/Vasool — 5-min walkthrough: https://youtube.com/watch?v=B0Iov6qAaqs
SentinelPrime — multi-threaded options system on live WebSocket data. Releases the GIL via Numba (nogil=True), actor model over daemon threads isolating DB I/O and API calls, fixed-size ring buffers to kill allocation overhead under 24/7 operation.
https://github.com/sriramvarun0636/SentinelPrime
Looking for backend/infra/agent-systems work at pre-seed to Series A — founding engineer or engineer #2-5. Eval and agent-safety teams too. Standard loops are fine, and I'll also take a paid 2-3 week trial sprint on a real bottleneck in your codebase.
I build the systems plumbing, concurrent pipelines, and secure runtime isolation that AI coding-agent startups are racing to ship. Happy to prove it directly on your codebase first — I'll take on a 2-3 week paid trial sprint on a real backend bottleneck before any longer-term conversation.
• AutoPatch-AI (https://github.com/sriramvarun0636/AutoPatch-AI): Autonomous, self-healing code remediation agent built on a LangGraph state machine. Engineered a zero-trust Docker execution layer (network_disabled=True, 256 MB RAM / 128 PID cgroup hard limits), streamed real-time telemetry over thread-safe queues via SSE, and built an AST-based line-level parser (cutting context/generation bloat by ~80%). Achieved a 100% resolution rate (5/5) on a custom Micro-SWE-bench harness with an average time-to-resolution of ~205 seconds.
• SentinelPrime (https://github.com/sriramvarun0636/SentinelPrime): Multi-threaded, regime-aware options trading system for live WebSocket market data. Releases the CPython GIL during numerical computations using Numba JIT compilation (nogil=True) and an Actor Model architecture over daemon threads to isolate DB I/O and API calls. Built thread-safe O(1) queues and fixed-size circular memory buffers to eliminate allocation overhead under 24/7 continuous operation.
• HydraPrime (https://github.com/sriramvarun0636/HydraPrime): Systematic XAUUSD strategy in MQL5 (C++ derivative), developed and iterated solo over two years — OTE/Fair-Value-Gap structural entries with ATR-adaptive sizing and multi-layered drawdown circuit breakers. Backtested across 96M+ real ticks over a single two-year window (2024–2026) that included a strong gold bull run: 1.37 profit factor, 13.8% max drawdown, 252% net return. Not yet forward-tested live, so I hold that return number loosely until it survives a different regime.
What I'm looking for:
Backend/infra/agent-systems roles at early-stage teams, pre-seed to Series A — founding engineer or engineer #2–5, wherever you're at. Standard interview loops work fine too, not just trial sprints.
I build the systems plumbing, concurrent pipelines, and secure runtime isolation that AI coding-agent startups are racing to ship.
AutoPatch-AI: An autonomous, self-healing code remediation agent built on a LangGraph state machine. I engineered a zero-trust Docker execution layer (network_disabled=True, 256 MB RAM / 128 PID hard limits), streamed real-time telemetry over lock-free thread queues via SSE, and moved from full-file diffs to Search/Replace block patching, cutting generation overhead by ~80%. Achieved a 100% resolution rate (5/5) on a custom Micro-SWE-bench harness with an average time-to-resolution of ~205 seconds.
GitHub: https://github.com/sriramvarun0636/AutoPatch-AI
SentinelPrime: A multi-threaded, regime-aware options trading system for live WebSocket market data. Built using an Actor Model architecture over daemon threads to isolate database I/O and broker API calls, with an asynchronous Token Bucket rate limiter and atomic SQLite (WAL) persistence.
GitHub: https://github.com/sriramvarun0636/SentinelPrime
HydraPrime: A systematic XAUUSD trading EA in MQL5. Backtested (not live) over two years on 96M+ real ticks, achieving a 1.37 profit factor, 13.8% maximum drawdown, and 252% net return.
GitHub: https://github.com/sriramvarun0636/HydraPrime
B.E Computer Science undergrad at NIT Allahabad (Class of 2027).
Looking for a founding or early-stage software engineering role on an AI-agent, developer tool, or high-performance backend systems team. Open to full-time or contract opportunities.
Remote: Yes (2pm-11pm IST — full EU hours + US-East mornings)
Willing to relocate: Yes
Available: Remote immediately, on-site from Dec 2026 — final-year B.Tech CS, MNNIT Allahabad
Technologies: Python, C/C++, Docker SDK (sandboxing, cgroups), LangGraph, FastAPI, WebSockets, Numba, SQLite (WAL), Linux
Resume: https://drive.google.com/file/d/1XXZ-d2hf8wEJYx0O-kX310qkeHr...
Email: sriramvarun636@gmail.com
GitHub: https://github.com/sriramvarun0636
Systems plumbing, concurrent pipelines, and secure runtime isolation for agent startups. The claim I'd rather be judged on: not that my agents are correct, but that being wrong is discoverable. Write-up: https://sriramvarun0636.github.io/
Vasool — compliance-gated payment recovery agent on Razorpay's test APIs. I pre-registered seven falsification criteria before running anything, and then lost against one: a dumb baseline that retries everything recovers 16.35pp more than mine. It also breaks policy in 1,000 of 1,000 seeded runs; mine breaks it in 0. That trade is the finding, and it's the second section of the README, not an appendix. The LLM never calls a tool — inert verdict type, no adapter to the execution plane, asserted by an import-graph test. Simulated outcomes, tagged as such in the source; every figure is a key in a committed manifest, so you can check any number without running anything. https://github.com/sriramvarun0636/Vasool — 5-min walkthrough: https://youtube.com/watch?v=B0Iov6qAaqs
AutoPatch-AI — autonomous code remediation agent on a LangGraph state machine. Zero-trust Docker execution layer (network_disabled, 256MB RAM, 128 PID cgroup caps), real-time telemetry over thread-safe queues via SSE, AST-based parsing that cut context bloat ~80%. https://github.com/sriramvarun0636/AutoPatch-AI — 90s sandbox demo: https://www.loom.com/share/104cf5ebcfc144a09f49c62830755408
SentinelPrime — multi-threaded options system on live WebSocket data. Releases the GIL via Numba (nogil=True), actor model over daemon threads isolating DB I/O and API calls, fixed-size ring buffers to kill allocation overhead under 24/7 operation. https://github.com/sriramvarun0636/SentinelPrime
Looking for backend/infra/agent-systems work at pre-seed to Series A — founding engineer or engineer #2-5. Eval and agent-safety teams too. Standard loops are fine, and I'll also take a paid 2-3 week trial sprint on a real bottleneck in your codebase.
reply