cs + math @ Penn State, Schreyer Honors College. 4.0 GPA, graduating December 2026. building cool stuff out of New York.

ICPC team @ Penn State. PennApps winner (85+ teams). HackPSU winner. UKMT Senior Math Challenge gold (top 6% nationwide). researching LLM agent security and RAG for political science codebook annotation.

open to work: fall 2026 internships and full time roles starting January 2027. open to relocating, on F-1 with CPT/OPT for now.

interested to work anywhere where the work is hard and shipped fast.

linkedin / x / viraj.mishra.81@gmail.com

most of what's below is live. click around.

• 📈 polymath - Bloomberg terminal for prediction markets. live Polymarket orderbooks, arbitrage scanner ranked by liquidity-weighted spread, 500-path Monte Carlo sims, portfolio VaR, orderbook-walking slippage in bps. keyboard driven. live

• 🧳 wander agent - MCP server that turns Claude into a travel agent with memory. 60 tools, zero API keys: flights, hidden city fares, visas, advisories, weather. tested on Python 3.10-3.12 in CI. pip install wander-agent. demo

• 🔭 bench - observability platform for AI agents. wrap your agent in one call: live dashboard, LLM eval scores, failure clustering, README badge, public leaderboard. npm + PyPI SDKs, Cloudflare Workers + Durable Objects. live

  

• 🚨 crisgo - crisis navigation built at Columbia's hackathon. 5 LangChain agents score incident credibility, Dijkstra routing with 600m penalty zones steers you around danger. AR camera nav + voice incident reporting. live

• 🎳 jarvis air bowl - bowling you play by swinging your arm at your webcam. MediaPipe tracks the swing, hand velocity history becomes hook spin and axis tilt, Cannon.js takes it from there. live

• ⚽ worldcup-forecaster - Dixon-Coles Poisson model fit on 13,779 international matches with time-decay weighting, Transfermarkt squad-value corrections, and xG substitution for WC 2022. predictions locked to git 60 minutes before kickoff and never edited. 10k Monte Carlo bracket sims updated live. 62.5% W/D/L accuracy on WC 2026 so far. dashboard
• 🦖 dino trainer - double DQN agent for Chrome dino, trained from a 9-value state vector. curriculum learning, behavior cloning warm start, replay memory. avg score 167.95 over 20 eval episodes, which is better than me. demo
• 🩺 medtrace - clinical decision support over 15,000 PMC-Patients records. hybrid BM25 + FAISS retrieval with reciprocal rank fusion, 13-stage pipeline, all inference local so patient data never leaves the machine. Palantir build challenge. demo
• 📚 studybuddy - upload a pdf, ask it questions or generate a quiz. chunking, embeddings, FAISS similarity search, Gemini. won best Streamlit + AI integration at HackPSU

• 🟩 wordle signal - Chrome extension that solves Wordle with Shannon entropy, feedback patterns encoded as base-3 integers. also solves Sedecordle (16 boards at once) with a hybrid early/endgame policy. demo
• 🐍 kinsnake - snake you steer by pointing your finger. React over WebSocket to FastAPI, which talks to a C process over named pipes. built with friends for a systems class
• 🌱 sustain-a-prompt - paste an LLM prompt, see what it costs in energy, CO2 and water, get a leaner version back. cuts token usage by ~40%. won MLH's Best Use of Auth0 at PennApps. live

💼 SIG Group - software and AI engineer intern, June 2025 to August 2025

• built a scheduling platform with React and Flask that replaced Excel, cutting planning time from 3 hours to 30 minutes
• deployed a RAG pipeline (Azure Search) over internal docs, so 130 engineers resolve queries instantly
• fine tuned YOLOv8 for defect detection on Azure ML, hitting 84% precision with only 100 training samples

🏫 Penn State - teaching assistant, August 2025 to present

• leading labs for 380+ students in digital logic design: Boolean algebra, FSMs, Verilog HDL

the course projects live in private repos, but ask me about any of these:

• ⚙️ 5-stage pipelined MIPS CPU - 32-bit RISC processor in Verilog. full 5-stage pipeline with custom hazard detection and forwarding units so data hazards don't stall execution. CMPEN 331 final
• 🧠 dynamic memory allocator - malloc, free and realloc from scratch in C. implicit free list with boundary tags, then an explicit LIFO free list so find-fit only scans free blocks, then footer optimization with a prev_alloc bit packed into headers so coalescing works without footers. CMPSC 473
• 🗺️ x86-64 kernel virtual memory - built a working kernel in C: full 4-level page table hierarchy (PML4E through PTE) in raw memory, identity-mapped 4GB for the kernel, separate user-space mapping, loaded CR3 myself, syscall dispatch from scratch. CMPSC 473
• 🔀 Go-style buffered channels in C - blocking and non-blocking send/receive plus a channel_select that multiplexes n channels at once, built on Pthreads mutexes, condition variables and a shared semaphore. CMPSC 473
• 💬 mycord - multithreaded terminal messenger in C. concurrent Pthreads server, non-blocking TUI straight from termios and ioctl. CMPSC 311 final
• 📐 conformal prediction for regression - ICP, CQR, jackknife and jackknife+ from scratch in Python. all four hit the 90% coverage target on heteroscedastic insurance data with zero distributional assumptions, CQR's adaptive intervals beat ICP's fixed-width bands, validated against MAPIE. MATH 415 honors project. colab

• 🛡️ llm agent security - benchmarking prompt injection attacks across small Qwen models with InjecAgent, 17 attack tools and 197 test cases, under Dr. Suman Saha. building a Task/Guard dual-model architecture that decouples instruction from execution, with a stateless tool-call authorization layer so untrusted external content can't trigger tool calls
• 🗳️ RAG for political science annotation - testing whether RAG-based codebook retrieval improves zero-shot LLM annotation, extending Halterman & Keith (2025), under Dr. Qunhua Li and Dr. Cassandra Tai. automating annotation of 7,236 qualitative records with a FAISS pipeline, all-mpnet-base-v2 embeddings and Mistral 7B across three datasets (BFRS, CCC, Manifestos). 8 ablations across retrieval depth, prompt structure and model; best setup improves F1 by 0.175. presented at Penn State's undergraduate research conference

this section rewrites itself daily from worldcup-forecaster's live numbers.

the model's title favorites right now: 🇪🇸 Spain 34.5% · 🇦🇷 Argentina 17.2% · 🏴󠁧󠁢󠁥󠁮󠁧󠁿 England 16.7%

live track record: 62.5% W/D/L accuracy (60/96 scored matches, random guessing gets 33.3%). every prediction locked to git before kickoff. full table + every scoreline

one shared board a day, six guesses, anyone can play. open an issue with your guess and a github action scores it, updates this board, and tells you what my entropy solver would have played instead.

⬜⬜⬜⬜⬜ _ _ _ _ _
⬜⬜⬜⬜⬜ _ _ _ _ _
⬜⬜⬜⬜⬜ _ _ _ _ _
⬜⬜⬜⬜⬜ _ _ _ _ _
⬜⬜⬜⬜⬜ _ _ _ _ _
⬜⬜⬜⬜⬜ _ _ _ _ _

2315 possible words remain. board resets at midnight UTC. yesterday's word: DRIED (unsolved)

working on research and answering my email.

if you're hiring interns for fall 2026 or new grads for jan 2027 fastest way to reach me is viraj.mishra.81@gmail.com. i reply within a day.