Open to AI engineering & SDE roles · graduating 2027
AI Engineer
Final-year Computer Science (IoT) at VIT Vellore. I build and evaluate LLM systems: defending agent memory from poisoning, eval pipelines that gate releases, and on-device AI in the browser.
$ now building Thymus
Work
Education
B.Tech in Computer Science and Engineering, specialization in Internet of Things
Schooling: St. Xavier's School, Jaipur
Toolkit
PythonTypeScriptJavaScriptC++20SQL
LLM agentsLLM evaluationLLM-as-judgeAgent memoryRAGEmbeddingsOllamaGemini Nano
Next.jsNode.jsPostgreSQLPrismaPreactChrome MV3CloudflareVercel
Power AppsPower AutomateSharePointESP32
Selected work
A defence against memory poisoning in LLM agents: it accepts real fact updates from untrusted channels while rejecting false ones shaped exactly like them. Benchmark of 491 matched pairs across 90 fact families in English, Hindi and Hinglish, including UPI, KYC and utility scams; five published defences reimplemented in one harness and run against agents from four model families.
An A/B evaluation pipeline for AI agents: graders, LLM-as-judge, failure analysis and a release gate in CI. On its own demo benchmark, V2 moved the pass rate from 28% to 96% at about 2.9× the token cost, and the report shows both.
Eight reusable Agent Skills for engineering, data and product work in the open SKILL.md format, all validated against the spec, with 80 evaluation cases and 8 end-to-end workflows checked in CI.
Finds open-source issues you can actually take: shows on GitHub whether a project is alive, how fast maintainers reply, and whether an issue is already claimed. Thread summaries run on-device with Chrome's built-in Gemini Nano. Published on the Chrome Web Store.
A single-node exchange core: price-time-priority order book, matching engine, pre-trade risk and kill switch, and a write-ahead journal with deterministic replay. 383 tests, and the same engine runs live in the browser.
Concurrency-safe seat holds, idempotent checkout with QR tickets, FIFO waitlists that auto-offer released seats, and an organiser dashboard with live revenue analytics. Tested with Vitest.
Also built
Live on this page
Two small things that run right here. Place an order against a live book, or look at a year of commits drawn the same way as the globe.
tape
A small JavaScript model of the idea. Trip the kill switch: new orders are refused but cancels still work, so the book drains. The real engine is C++20 with 383 tests, a write-ahead journal and deterministic replay. Run the real one ↗
Hover a day.
© 2026 Safal Gupta · safalgupta.tech · press 1–6 to jump between sections
↑↓ move ↵ open esc close