Building systems across ML, hardware, finance, and autonomous agents.

🎓 2nd-Year B.Tech – Electronics & Communication Engineering (KIET Deemed to be University)
🏆 3rd Place – Agentic AI Hackathon @ MAIT (1,100+ registrations)
🥈 Rank 53 – Partcl × HRT Macro Placement Challenge (1.5286 avg HPWL, 17 IBM benchmarks)
📊 IMC Prosperity 4 – Finals ranked #660 overall, #64 manual, #22 country (4.65L XIRECX)

• Building multi-agent AI systems (ReAct framework, LLaMA-3.3-70B, Groq inference)
• Quant trading algos (Black-Scholes, market making, premium selling strategies, pure-Python implementations)
• Hardware & placement optimization (simulated annealing, edge extraction, parallel restarts)
• Competitive programming — solving across DSA, system design, and optimization problems
• Full-stack development — Node.js/Express, React, FastAPI, Docker, Kubernetes

📧 pulkit300405@gmail.com | 🔗 LinkedIn

3rd Place – Agentic AI Hackathon @ MAIT | ReAct framework with autonomous buyer/seller/mediator agents

• Built 3 autonomous LLM agents (Buyer, Seller, Mediator) negotiating deals in real-time using ReAct (Reasoning + Acting) framework
• Integrated Groq LLaMA-3.3-70B for fast agent inference (~500ms/turn)
• Implemented ZOPA calculator (Zone of Possible Agreement) — mathematically enforces Pareto-optimal agreement bounds
• Sentiment scoring on each agent message + structured negotiation history tracking
• Prompt injection defense — validates agent prompts before execution to prevent manipulation
• Result: Agents converge to Pareto-optimal agreements in 3-4 rounds with 80%+ convergence, 7.3/10 avg sentiment

Node.js Express React Groq API LLaMA-3.3-70B ReAct Framework Vercel

Multi-turn RL environment for fraud investigation — agents reason under uncertainty, gather evidence, issue verdicts

• Designed OpenEnv-compatible RL environment where agents investigate sessions via 5 signal types (IP velocity, device fingerprint, login frequency, geo anomaly, request patterns)
• Agents must balance investigation cost vs. verdict confidence — max 3-8 steps per session depending on difficulty tier
• Deterministic heuristic grading (no LLM-as-judge) ensures reproducibility and speed
• Multi-difficulty tasks: Easy (obvious fraud), Medium (mixed signals), Hard (adversarial evasion)
• Baseline performance: Qwen-72B scores 1.85 (easy), 1.10 (medium), 0.75 (hard)

Python FastAPI PyTorch OpenEnv Docker HuggingFace Spaces

Distributed load testing platform for trading engine evaluation — 5K+ concurrent bots, latency p-percentiles, correctness validation

• Engineered Bot Fleet (Go + goroutines) generating 5K+ concurrent connections with realistic order patterns
• Built Submission Handler — accepts contestant code, containerizes in Docker, runs in isolation
• Telemetry Ingester — captures latency (p50/p90/p99), throughput (TPS), validates correctness (FIFO, no double-fills)
• Real-time React Leaderboard with WebSocket updates
• Full stack: Microservices (Go), PostgreSQL + TimescaleDB (metrics), Kubernetes-ready infrastructure

Go Docker Kubernetes PostgreSQL TimescaleDB React Microservices

Multi-strategy quantitative trading bot — market making, momentum, options pricing, finals rank #660

• Implemented Black-Scholes option pricing with pure-Python normal CDF (no external libs)
• EMA momentum strategy for directional bets + market making for liquidity provision
• Premium selling on OTM vouchers — delta hedging + Greeks calculation
• Competed as "KENSAI TRADING" (2-person team) — achieved #22 country rank, #64 manual rank
• Total PnL: ~4.65L XIRECs across 5 rounds

Python (stdlib only) Black-Scholes Greeks Market Making Momentum Trading

Simulated annealing + coordinate descent for circuit macro placement — 17 IBM ICCAD04 benchmarks

• Implemented SANetPlacer using simulated annealing with tuned parameters (T_start = max(cw,ch)*0.25, T_end = max(cw,ch)*0.0008)
• Edge extraction via _build_edges_from_net_nodes() for NG45 compatibility
• Refinement phase (3000 iterations) — coordinate descent + pairwise swaps
• Result: 1.5286 avg HPWL (0 overlaps) across all benchmarks, below RePlAce baseline (1.4578)
• Key insight: Bottleneck is eval speed, not search strategy — optimized for fast evaluation

Python Simulated Annealing Coordinate Descent Parallel Restarts Circuit Design

• 3rd Place — Agentic AI Hackathon @ MAIT (March 2026, 1,100+ registrations)
• Rank 53 — Partcl × HRT Macro Placement Challenge (May 2026)
• IMC Prosperity 4 Finals — #660 overall, #64 manual strategy, #22 country (April 2026)
• FOSSEE eSim Fellowship — Task 5 submission (Tool Manager), IIT Bombay
• Competitive Programming — 300+ DSA problems on LeetCode/GFG

Always learning. Open to AI/ML internships, quant finance roles, hardware optimization challenges, and hackathons.

⭐ Check out the projects — feedback welcome!