| [ Web Proxy ] |
| Viewing: https://paramsc.dev/ | [Back] [Original] |
AI Systems & Full-Stack Engineer.
MS in Computer Science at UNC Chapel Hill.
UNC Chapel Hill CS Dept
UNC Chapel Hill CS Dept
Information Technology Services, UNC Chapel Hill
DRDO, Bangalore
Bharat Blockchain Network
Palo Alto Networks
Amazon Web Services
Capability-based security sidecar in Rust. Replaces over-privileged API keys with task-scoped, short-lived sessions for AI agents stateless authorization, K8s-deployable Docker container, Biscuit token facts. Companion MCP server + Python SDK exposes security tools to agents in LangGraph.
Real-time observability platform on AWS Next.js + React + FastAPI. OpenTelemetry instrumentation for AI agents, nested checkpoints to SQLite for time-travel debugging. Handles 50+ concurrent workflows under 20 ms.
Custom INT8 MAC instructions extending the CVA6 core via CVXIF. Attention microkernel with deterministic loop execution removes CPU dispatch overhead 7.59 p99 latency cut on batch-1 attention (307 34 cycles).
$0/month idle, sub-50 ms cold starts. gRPC ingestion gateway, Upstash Redis semantic caching, Upstash Vector embeddings 3060% LLM token reduction. Playwright e2e for UI reliability.
RAG chatbot for paramsc.dev, indexed over Param's CV, resume, projects, OSS PRs, writing, education, and practice logs. Uses build-time local MiniLM embeddings, bundled JSON vector search, Groq Llama 3.3 streaming, and citation chips with a $0/month operating target.
Hybrid TLA+ KLEE framework. Model-checker counterexamples drive symbolic execution; property-directed loop auto-generates invariants and KLEE harnesses. Verified safety properties for Raft.
FastAPI + Python pipeline. Semantic search resolves common queries instantly from an FAQ knowledge base; React frontend escalates the rest. Serves 1,000+ users daily.
Published to VS Code + Cursor marketplaces. PyMuPDF + python-docx engine for high-fidelity text extraction. Auto-detection layer lets AI agents read local docs for context-aware code generation.
Python + LangGraph router for agentic queries between local and remote MCP tools. Semantic routing enforces data boundaries while keeping sub-second latency under load.
BERT-based classification feeding a ChromaDB vector store. Custom cross-encoder reranker, grid-searched hyperparameters high-precision document retrieval.
Resolved race conditions in parallel tool execution. Co-architected a unified citation interface (8.5k+ LoC) that standardizes source attribution across OpenAI, Anthropic, and Gemini providers.
INT8 quantization checks + automatic FP8 fallback for SM100+ kernel loading. Keeps the runtime stable when newer GPUs lack expected kernels.
Diagnostic suite + connection-pool fix to prevent leaks when SSE streams fail mid-flight.
Corrected how RV64 interrupt causes are recorded; eliminated unintended latch inference. Improves cross-platform synthesis on a CPU that ships in real silicon.
Fixed a YAML parsing bug in skill-creator template (bracketed descriptions misread as lists). Sped up bbox validation 3 by grouping checks per page to skip redundant cross-page comparisons.
Multi-strategy capture (CDP, Firefox-native) handles pages that overflow the viewport without manual scroll-stitching.
Cut redundant browser launches; added respectful rate limiting with strict per-source article caps.
Im not a professional mathematician. Im someone who loves algorithms, proofs, and the moments when pure math collides with the real world. So when OpenAI announced their internal model had
A quick heads-up: This post is a bit of a deep dive. While I usually aim for brevity, the history of world models, stretching from 1943 to todays multi-billion dollar research labs, is far too
DRDO, Bangalore
Google Developer Group, Pune
Blockaders Club, PES Modern College
Want to chat? Reach out anytime use the email card or links below; your mail app opens in one tap.
| Web Proxy Viewer | New URL | Original Page |