Michelle Feng · San Francisco
I build AI memory at Iditor. The product remembers your conversations with an AI, so the next one can pick up where you left off. On this site I build the same kind of tools for my own life:
Founding engineer at Iditor · Boston University CS ’25 · 4 published agent skills
2024 → now
Everything I build does one thing: it takes a messy pile of signals and turns them into one clear next step. So far that has taken three forms.
Complete systems, owned end to end. An energy platform that answers one question: after everything I already spent today, what is still safe to take on? Research with real numbers: our debiasing method raised female representation in generated images from 29% to 44.5%, and our RL method converged 10× faster. At Iditor, EchoChat went live with chat, multimodel inference, memory, and retrieval.
What this proves: I can own a system from database to interface and ship it live.
Three notebooks in two years, each distilled from the last. The projects change; the method carries over.
Don’t take the page’s word for it. Install one of my published agent skills and judge the craft directly:
Prefer to browse? github.com/myfeng10 ↗ · or try the working prototypes live: Energy, Identity, the lottery.
A system I can’t make someone else understand isn’t finished. Most of my work is the interface between a complicated process and a person who needs to stay oriented inside it.
See the visualization archive →Big ideas earn adoption by shrinking. I turned a whole platform into four installable skills. I ship the version people actually use.
See what I’ve been building →Numbers where they exist, artifacts where they don’t. Every claim on this page links to something you can open, read, or run.
See the install command →Each project becomes material for the next form of the same idea. The archive below is that process, kept honest in public.
See the timeline →Field notes · 手记
I keep a public journal on Rednote (马歇不歇 ↗). It has thirty-some entries on how I talk with people, how I manage my energy, and how work gets unstuck. A few of them, translated from the originals:
原来 coffee chat 不一定要“表现得很好”
What made coffee chats work was giving the other person, in the shortest possible time, a true picture of who I am and where I’m stuck, so their experience could actually land on my problem. Real beats polished: when I speak honestly, people slow down and share the specific version of their story. When I sound official, they answer officially.
「在表达中，真实比“完美”更有力量。」
原帖 · original ↗最近才搞懂为什么会沟通时内耗
My problem was never technique. I would speak without deciding what the sentence was for. Now I sort first. Some words only need to mark that I’m present and my position hasn’t moved. Saying them completes the task, whatever the reaction. Other words are meant to move things forward. There, feedback matters as a signal of where information hasn’t aligned yet.
「先判断沟通目的，再决定我需不需要在乎反馈。」
原帖 · original ↗对方还没进入状态，我已经讲完了
I used to open with all the background, every premise, every worry. I thought that was being responsible. But before the other person is ready, more information just loses them faster. The practice now: say the one sentence I care about most. Then stop. If they’re interested, they’ll ask.
「先只说那一句我现在最在意、最核心的点。说完就停一下。」
原帖 · original ↗希望这个世界再没有拖延
Deadlines work because they make things simple: the task’s weight becomes concrete, the open paths collapse to one, and “perfect” quietly becomes “done, then adjust.” So I shrink decision complexity on purpose: one deliverable, one start time, permission for the first version to be rough.
「我们以为自己需要压力才能行动，其实需要的是更低的决策复杂度。」
原帖 · original ↗工作之后发现，努力不再能解释一切了
In school, effort explained outcomes. At work, whether something moves depends on whether it sits at the same priority for everyone involved. When a thing stalls now, I check whether the expectations, the standard of done, and the handoffs were ever actually aligned.
「很多矛盾不是靠情绪解决的，而是靠把问题讲清楚解决的。」
原帖 · original ↗有时候，是环境在决定你怎么活
A lot of what we call personality or discipline is state. Swap the environment and the behavior evaporates. So the real work happens in the good states: the connections you keep and the attempts you make are exits you’re leaving for a future self who won’t have the strength to dig one. When the low phase comes, those outside forces pull you out.
「为“未来可能状态不好的自己”，留下出口。」
原帖 · original ↗Daily practice
One source, one tiny task, one sentence, drawn fresh each day. Built to keep interview prep light and consistent.
Archive
The short version is above. This is the long one: how each phase actually felt, what it taught me, and why it mattered. If the top of the page earned your attention, this part is for you.
Making inner systems legible
This period concentrates on memory, direction, energy, identity, and the question of how to turn lived patterns into surfaces I can actually use later.
A set of installable agent skills that turn energy, identity, weekly review, and archive practices into reusable decision workflows.
The shift is from personal ideology to portable behavior: the agent can notice patterns, make decisions, and hand back one next move.
This site became a project about preserving meaning across time, not just displaying outputs. The redesign started when I realized the old structure could describe me, but could not help me remember myself.
The site needs enough structure to preserve memory, but not so much structure that it produces fake completeness.
A tiny walking experiment about direction, drift, and the strange gap between what the body thinks it is doing and what it actually does.
The original mood was curiosity, not fear. It should stay playful, sleek, and quietly strange.
A working system for translating sleep and the structure of a day into guidance about what kind of work is still safe to do.
The real question underneath it is not “how productive am I?” but “what kind of thinking is safe now?”
A system that reads recurring actions as identity votes, trying to make becoming legible before it hardens into a story I tell too late.
The interesting part is not the label itself. It is the accumulation of repeated behavior over time.
The bridge project: one place to capture a day, read energy and identity together, and keep a running history instead of isolated analyses.
This is where multiple tools started wanting to become one ritual.
A running note system for tracking patterns, shifts, and growth in a way that stays close to actual life instead of polished summaries.
This is the record underneath the cleaner interfaces.
Threshold year
Graduation and startup life changed the pace, stakes, and texture of work. This is the environment that later archive projects grow out of.
A transition from school structure into startup velocity, with a corresponding shift in how work, time, and responsibility felt day to day.
This period is less about one artifact than about the environment that changed what kinds of projects became necessary.
Full-stack work across chat systems, multimodel inference, memory, and retrieval, in the context of a live social AI product.
This is where product engineering, memory systems, and real user-facing constraints met each other.
Research, teaching, and explanation
Before the recent personal systems work, the same instinct was already present in research, teaching, and interactive explanations.
Research, teaching, and explanation work kept reinforcing the same instinct: complicated systems matter more when someone else can actually follow them.
This thread explains why so many later projects turn invisible processes into readable surfaces.
Older competitive programming and visualization work that still shows a consistent preference for explaining structure through interaction.
These pieces are earlier, but they still belong because they show a stable way of thinking.
I wrote this page for the people who read this far and recognized something: a way of working, a question they also can’t put down, a younger version of themselves. I’m still looking for my limits. The fastest way I know to find them is working beside people who already found theirs. Everything above is my evidence that the investment compounds.
— Michelle 马歇不歇