Building ether.fi Cash, a non-custodial Visa crypto card on liquid-restaking infrastructure where users spend ETH/USDC while their assets continue earning restaking yield.

• Own real-time settlement bridging Visa rails and on-chain vaults (authorization, capture, reconciliation).
• Built the Borrow Mode backend - USDC borrowing against eETH collateral via lending-pool integration.
• Designed an event-driven cashback pipeline (wETH rewards) on Kafka.
• Hardened KYC/AML and risk monitoring with chain-analytics providers; extended multi-chain support across Ethereum L1 and L2s.

$ Go · Solidity · Ethereum · EigenLayer · Kafka · PostgreSQL · Redis

Worked across traditional cross-border payments and Visa's stablecoin / crypto initiative - the bridge between card networks and digital assets.

• Led an 8-engineer team on a real-time payments platform (monolith plus eight microservices) sustaining 2,000+ TPS at 99.98% uptime.
• Delivered Treasury as a Service for the APAC hub - 15 currencies, clients in 60+ countries, powering roughly 80% of APAC liquidity flows through Visa Direct A2A.
• Built the Network Validations Framework - configurable payer/beneficiary identity checks across 100+ SWIFT message types.
• Prototyped on-chain payment settlement in Solidity and Solana as part of the crypto/stablecoin track.

$ Java · Spring Boot · Solidity · Rust · RabbitMQ · Hazelcast · Docker · Kubernetes

Joined as an intern during the final years of my degree at Delhi College of Engineering and grew into a founding member of a ~7-person team building a multi-chain liquid-staking protocol.

• Built Rust / CosmWasm smart contracts powering liquidity pools that scaled past $500M+ in staked assets across multiple chains.
• Led gas-fee analysis and optimization - benchmarking storage and swap costs on Terra/CosmWasm to keep contract execution economical at scale.
• Built a validator analytics platform monitoring 1M+ nodes (React, MongoDB, Ethereum APIs) and APR dashboards.
• Designed cross-chain architecture and on-chain/off-chain synchronization pipelines.

$ Rust · CosmWasm · TypeScript · Python · React · MongoDB

Cache-Aware Client-Side Request Planning for Black-Box LLM APIs. When you consume an LLM through a paid per-token API you pay for the tokens the server processes, so the only lossless lever a client has is re-ordering requests to hit the provider's prompt cache. Formalizes that design space and ships a greedy prefix-clustering scheduler that shapes request order to maximize cache hits - up to 60% billed-cost reduction on an agentic workload at zero quality loss, Pareto-dominating prompt compression and semantic caching.

DOI 10.5281/zenodo.21386594 · pjdurden/cache-aware-request-planning

Perplexity Holds, Programs Break: Executable Correctness as a Blind Spot of KV-Cache Compression. KV-cache compression is benchmarked almost entirely on token-overlap and retrieval metrics that never check whether generated code actually runs or a tool call is schema-valid. Introduces kv-exec-bench, an open benchmark measuring code unit-test pass@1 and tool-call JSON-Schema validity under compression, built on NVIDIA's kvpress so any press works unmodified.

DOI 10.5281/zenodo.20805562 · pjdurden/kv-exec-bench

Straggler-Aware Elastic Membership for Decentralized Training. A zero-GPU discrete-event simulator of DiLoCo-style decentralized ML training plus a straggler-aware membership policy for slow-but-alive nodes that existing decentralized-training stacks don't handle. 4.59x faster than the baseline on a persistent-straggler scenario, validated against a torch.distributed/gloo DiLoCo loop.

· pjdurden/churn

The standard high-throughput LLM inference & serving engine - build-correctness fix in the precompiled-flag test suite (#44942).

Rust search engine - fixed a ranking-rules ordering bug that silently dropped matching hits (#6437).

High-throughput LLM/VLM serving engine - fixed the prefill/decode router's cache-aware routing keying chat requests on the first message only; routing on the full conversation lifted KV-cache hits from ~69% to ~96% and output throughput from ~678 to ~1078 TPS (#27430).

Hugging Face's minimalist Rust ML framework - Qwen3 produced incorrect output for any batch size > 1: the causal mask allocated a batch-independent buffer but claimed a (b, 1, tgt, tgt + offset) shape, so batch row 0 read the correct mask and every row after it read past the buffer; fixed by shaping the mask (1, 1, ...) and letting the existing broadcast_add apply it across the batch (#3586).

The leading Go implementation of the Model Context Protocol (MCP) - fixed getServerMetadata returning (nil, nil) instead of an error, closing a silent-failure path in OAuth discovery (#904).

The Rust LLM inference & serving engine: fixed reversed FCFS priority in the PagedAttentionScheduler preemption path so the oldest request is preempted last (#2250), and validated GGUF special-token ids against the vocab to prevent an out-of-bounds panic on model load (#2282).

The OP Stack monorepo powering Ethereum L2s (Base, OP Mainnet) - fixed an op-wheel metrics bug that wrote block gas twice and left the base-fee gauge unset (#21127).

The KV-cache store & transfer engine behind Kimi, used as a disaggregated KV backend by vLLM and SGLang - mooncake_master bound its RPC and HTTP servers to the numeric wildcard 0.0.0.0, which the acceptors resolved through getaddrinfo(), so environments that answer EAI_NONAME for numeric literals killed startup with bad address: 0.0.0.0; fixed by pinning the dependency to a revision that parses numeric IP literals directly (#2919), and fixed an SSD-offload duplicate-key storm under concurrency: when two offload flows shared a KV prefix block, the bucket backend's intentional single-writer-per-key OBJECT_ALREADY_EXISTS rejection was treated as fatal by FileStorage::OffloadObjects, aborting the whole offload and leaving the decode node with INVALID_KEY floods; made duplicate-key rejection a recoverable per-bucket condition (#2967).

The vLLM project's Kubernetes-native LLM-serving control plane - fixed the ZMQ KV-event decoder dropping group_idx/medium/lora_name from vLLM's BlockStored event, which caused false prefix-cache matches on hybrid-attention models (#2384), and made the KV-event indexer purge a pod's cached prefixes on AllBlocksCleared so evicted blocks are no longer served as stale cache hits (#2385).

The Rust EVM behind Foundry & reth - removed a const-eval panic path in the stack interpreter (#3735).

The Envoy-based gateway for AI/LLM traffic - the MCP proxy failed initialize with a 500 when a backend's initialize SSE response opened with a keep-alive / empty data: event before the JSON-RPC result (seen on some FastMCP backends); the SSE parser now skips non-response events instead of treating them as a fatal parse error (#2267), and the Anthropic translator dropped the input/cache token usage reported on message_delta, so it was left out of the final usage totals; fixed to merge it in (#2292).

The Go node for the Celestia data-availability layer - unified the header TestSuite constructors behind functional options (#5041).

The Rust Ethereum-types & sol! toolkit used across Foundry, reth and the wider Rust EVM ecosystem - the sol! macro silently dropped Debug/PartialEq/Eq/Hash derives on the event/error enums it generates for contracts with overloaded events (e.g. Uniswap V3's two Swap events), because synthetic _N-suffixed variant names didn't resolve; fixed it to compute derivability from the underlying parameter types (#1118).

The Rust Ethereum L1/L2 execution client by LambdaClass - replaced a too-broad datadir non-empty startup check with an actual-DB probe, so unrelated files (e.g. an EthDocker JWT secret) no longer block a fresh node from booting (#6786).

The constrained-decoding engine behind structured / JSON-Schema output in vLLM, SGLang and llama.cpp - fixed multipleOf rejecting negative multiples (#357) and made max_tokens=0 rules compile to the empty string (#356).

The Rust JSON-Schema to regex core behind Outlines' structured generation - the date format regex applied a uniform 01-31 day range regardless of month, so constrained decoding could emit impossible dates such as 2022-02-31 and 2022-04-31; made the day range month-aware, deliberately leaving leap years unvalidated to keep the compiled regex bounded (#258).

A canary for silent KV-cache-compression failures - functional accuracy (code execution / tool-calls) vs perplexity under quantization & token eviction, scored by a Silent Degradation Score. pjdurden/kv-canary

Deterministic discrete-event simulator of elastic-membership decentralized ML training, plus the straggler-aware membership policy behind the paper above. pjdurden/churn

Cycle-accurate 6502 CPU emulator down to bus/processor pins, addressing modes, and internal functions, with an OpenGL / Pixel Game Engine graphics frontend. pjdurden/friendly-neighbourhood-6502

DEX swap contract for LUNA/UST/CW20 tokens, with a gas-fee analysis tool for the Terra blockchain. pjdurden/CoinSwap

Benchmarking contract measuring gas costs of storage patterns on Terra/CosmWasm. pjdurden/Gas_analysis

Rust payment-settlement program deployed on Solana. pjdurden/Solana-PaymentSystem-Visa

React dashboard surfacing staking APR and validator analytics. pjdurden/Stader-Apr-Dashboard-

Real-time identity and account verification API for cross-border payments. pjdurden/veri5-mvp

Languages

Blockchain & Web3

Backend & Infrastructure

AI / ML Systems

Achieved Codeforces Expert during university (pre-2022, before AI coding assistants) - a credential earned through unassisted algorithmic problem-solving under contest conditions.

• Codeforces - Expert · pjdurden
• CodeChef - peak rating 1890 · pjdurden
• LeetCode - peak rating 1679 · Pjdope
• Code-A-Thon - Rank 5 (College Rank 1)
• Cleared the Facebook Hacker Cup Qualification Round (2020); Google Kick Start Round B (2021).
• Mentored 300+ students in algorithms and software engineering through GirlScript.