$ icat felix.png && whoami

Felix Lin

i like building reliable software — here's some of my work

$ ls links/

$ tree projects/ · click a row to cat it

├──shipped/production — other people depend on it

2026 zylosnode.js · agent infra › channel integrations, plugin system, and reliability infra for an open-source agent platform

$ cat projects/zylos.md

# zylos

what: an open-source platform that turns LLM runtimes (Claude Code, Codex) into persistent agents — memory across sessions, real communication channels, no human in the loop
comms: six production channel integrations — Discord, Microsoft Teams, WhatsApp Business, LINE, Zalo, Zalo Personal — each against one channel-conformance contract on a SQLite-backed gateway; 400+ users on deployed agents
hardening: webhook signature verification, SSRF-guarded media fetch, and timing-safe token comparison across all six, plus JWKS-validated Entra ID auth across four sovereign Azure clouds for eight registry components
onboarding: the dashboard connect flow for every channel (Next.js/TypeScript, AES-GCM credential encryption, WhatsApp Business OTP path) — 30% of all human-agent IM connections
plugins: the component system: modular plugins with hot-reload config, lifecycle hooks, and a public registry, enabling community integrations across 10+ channels
infra: production agent services with health monitoring, crash recovery, heartbeat probes, and context-aware session rotation
links: zylos-ai org ↗

2026 zylos-recallnode.js · retrieval › retrieval layer that surfaces the right memory into an agent’s live sessions

$ cat projects/zylos-recall.md

# zylos-recall

what: automatic recall for a persistent-memory agent — relevant memory injected as it works
problem: an agent can store months of memory but won't resurface it unless someone loads it by hand
approach: BM25 and dense-embedding retrieval fused by Reciprocal Rank Fusion, cross-encoder reranked every turn; fail-open behind a 1s timeout
outcome: gated on an 18-case golden set at nDCG@k 0.95 / injected-F1 0.85 — a regression threshold, not a benchmark. live against a production agent's memory; the eval side became mem-eval
links: repo ↗

2026 mem-evalpython · docker › docker harness that measures what LLM agents actually retain over long horizons

$ cat projects/mem-eval.md

# mem-eval

what: measures what agents actually retain — not what they claim to
problem: memory demos test recall with the answer still in context; that measures prompting, not memory
approach: a dockerized distill/answer/judge harness replays chat histories one memory-sync per simulated day, then probes the agent on its real runtime after controlled noise
outcome: 125 LongMemEval questions scored to a 75% retention baseline; traced the weakest category (20%) to upstream distillation rather than answer-time retrieval
links: repo ↗ · harness readme ↗

results — claude runtime, vanilla-memory arm · longmemeval-s · 125 questions · overall 75%

knowledge-update93%

single-session-user93%

temporal-reasoning82%

multi-session73%

single-session-assistant73%

single-session-preference20%

accuracy by question type · 15–50 questions per type · full run data in the repo under data/runs/

2021 gait-rehab-appflutter · dart › flutter app for stroke-rehab patients, built at Chapman’s gait lab

$ cat projects/gait-rehab-app.md

# gait-rehab-app

what: a Flutter/Dart app for stroke-rehab patients, built as a mobile-development intern at Chapman’s Gait Rehabilitation & Research Lab
problem: clinicians see a patient an hour a week and can’t tell whether the exercises happen in between
approach: rep tracking, progress charts, and adherence monitoring in one patient-facing app
outcome: gave clinicians visibility into between-visit adherence
links: internal to the lab; not public

├──research/measured and peer-reviewed

2022 smartmetercomputer vision · time series › leak detection from an analog water meter — computer vision reads the dial, seasonal ARIMA flags divergence; MLNLP '22

$ cat projects/smartmeter.md

# smartmeter

what: a latch-on accessory that turns an ordinary water meter into a leak detector
problem: household leaks hide inside normal usage, and analog meters can't flag them
approach: a Cloud Vision dial reader takes readings off an unmodified analog meter; seasonal ARIMA learns the household's daily rhythm from 15-minute readings
outcome: held-out one-day forecast validated by RMSE against a seasonal-naive baseline; divergence from the forecast flags a candidate leak. proof of concept on seven days of data — the flagging rule is the paper's design, not a deployed system
links: paper ↗

└──coursework/CMU · code private, academic integrity

2025 memory-allocatorc · systems › malloc/free in C — segregated free lists, footerless blocks, mini-blocks; 11,031 KOPS at 74.1% util

$ cat projects/memory-allocator.md

# memory-allocator

what: a general-purpose dynamic memory allocator (malloc/free) in C
problem: fast and low-waste pull against each other
approach: segregated explicit free lists with better-fit search, footerless allocated blocks, two-word mini-blocks for the smallest allocations, adjacent-block coalescing
outcome: 11,031 KOPS throughput at 74.1% memory utilization
links: code private (academic integrity)

2025 network-proxyc · systems › multithreaded HTTP proxy in C with an LRU object cache

$ cat projects/network-proxy.md

# network-proxy

what: an HTTP proxy in C that serves many clients at once
problem: repeated requests each cost a round-trip to the origin, and a single-threaded proxy stalls everyone behind one slow request
approach: detached thread-per-connection, with a mutex-guarded LRU object cache answering repeats
outcome: fewer redundant network calls, lower latency under load
links: code private (academic integrity)

2025 unix-shellc · systems › tiny unix shell in C — job control, signals, I/O redirection

$ cat projects/unix-shell.md

# unix-shell

what: a mini Unix shell (tsh) — runs programs and manages foreground/background jobs
problem: children exit and signals arrive whenever they want; naive handling loses jobs or deadlocks
approach: fork/exec with per-job process groups; SIGCHLD/SIGINT/SIGTSTP handlers; signal-blocked critical sections with a sigsuspend foreground wait; I/O redirection
outcome: no lost children, no zombie buildup, ctrl-c/ctrl-z land on the right job
links: code private (academic integrity)

2026 rl-pongpython · pytorch · rl › advantage actor-critic agent that learns atari pong from reward alone

$ cat projects/rl-pong.md

# rl-pong

what: a reinforcement-learning agent trained to play Atari Pong (ALE/Pong-v5, restricted to up/down)
problem: no labels — the agent only sees a score and has to work out which of thousands of frames mattered
approach: Advantage Actor-Critic in PyTorch — separate policy and value networks (2×256 hidden), n-step advantage returns, decoupled policy and value objectives
outcome: learns from reward alone; reported on reward curves, not a single accuracy number
links: code private (academic integrity)

$ cat now.txt

exploring agent memory — how AI systems retain, recall, and prove it. ask me about LLM agents, eval harnesses, or retrieval.