Ashesh Kaji

Coursework: System Optimization Methods, Computer Systems Architecture, Special Topics in Computer Engineering. Research focus: portfolio optimization as layered control, convex optimization, reinforcement learning for financial decision-making.

Honors thesis on statistical modeling of environmental exposures and neuroimaging data. Coursework: Deep Learning, Reinforcement Learning, Probability & Statistics, Linear Algebra, Algorithms.

• Architect and deploy production AI infrastructure: LLM serving pipelines, MCP (Model Context Protocol) servers, and agent orchestration systems with LangChain and AI-SDK
• Design custom data transformation and fine-tuning strategies across multi-modal data lifecycle requirements
• Build and maintain ACP-compliant agent communication protocols for distributed AI systems

• Led development of multiple production AI products: LLMs, small language models (SLMs), statistical mapping models, and semantic database memory retrieval systems
• Designed end-to-end data pipelines for fine-tuning and deploying models with rigorous evaluation methodology
• Built RAG systems integrating vector databases with serverless deployment on Azure and AWS

• Built and deployed a production RAG system for document intelligence using LLMs, vector databases, and serverless cloud infrastructure on Azure
• Designed data extraction pipelines with rapid versioning, architecture design, and production testing cycles

• Applied statistical modeling (regression, causal inference) to analyze UK BioBank and ABCD Study datasets for relationships between biomarkers and neuroimaging outcomes
• Developed reproducible analysis pipelines in Python using NumPy, Pandas, and Statsmodels for epidemiology research

End-to-end quantitative research pipeline for portfolio construction. Designed a walk-forward backtesting engine with convex quadratic programming (Markowitz with ℓ₁ turnover penalty), regime detection via statistical jump models, and a policy-gradient reinforcement learning router for dynamic strategy selection. Modeled multi-agent competition for liquidity as a Nash equilibrium problem using iterative best-response with crowding penalties. Validated across top-100, top-250, and top-500 equity universes with provenance-ledger tracking to eliminate look-ahead bias.
Tech: Python, cvxpy, NumPy/SciPy, PyTorch, PufferLib, Nautilus Trader, D3.js
Research presentation ↗  ·  Interactive D3 demo ↗  ·  github.com/ashesh8500/fractal ↗

Implemented a paper-parity hardware accelerator for the zkPHIRE SumCheck protocol on Xilinx PYNQ-Z2 (Zynq-7020). Designed Montgomery multiplier units to reduce DSP utilization from 200+ to 8-12 per multiplier. Built a fused update+extension pipeline with banked scratchpads for memory-efficient MLE evaluation. Optimized resource allocation to fit within 220 DSP and 280 BRAM constraints using Vitis HLS and SystemVerilog.
Tech: Vitis HLS, SystemVerilog, C++, Python, Vivado, FPGA (Xilinx Zynq-7020)
Independent research project — paper reference: zkPHIRE (HPCA 2026)

Deep reinforcement learning for autonomous driving using the CARLA simulator. Implemented and compared policy gradient methods, DQN variants, and sim-to-real transfer strategies with real-world sensor data integration.
Tech: Python, PyTorch, CARLA, OpenAI Gym
github.com/ashesh8500/fp185 ↗

Local-first media library pipeline: content-aware deduplication with SHA-256, hardware-accelerated transcoding to MP4, and batched cloud upload. Built in Rust with a terminal UI for large-scale photo/video library management.
Tech: Rust, FFmpeg, rclone, terminal UI
github.com/ashesh8500/mediasync ↗

Portfolio site with a browser-native AI assistant running a 1-bit quantized Bonsai-1.7B ONNX model on WebGPU. Built a Cloudflare Worker proxy for DeepSeek API with rate limiting and SSE streaming. Inference toggle between local WebGPU and cloud models.
Tech: JavaScript, ONNX Runtime Web, WebGPU, Cloudflare Workers, Transformers.js
asheshkaji.com ↗