I build production AI systems and side projects. Writing about what I learn in engineering, productivity, and figuring out life.
San Francisco, CA 🌉
Joined March 2020
- Tweets290
- Following242
- Followers272
- Likes262
Pinned Tweet
my version of small talk involves voice agents , chat bots and evals !
👋 I’m deepali! New to writing on X.
Outside of AI engineering, love getting creative with my dev workflow !
If that sounds like your kind of small talk? Let’s connect 🤝
19
23
537
there's a lot between getting a model to answer and building an app around it. context, state, deciding what code should control.
i put those pieces into an illustrated guide for software engineers building their first ai product:
deepalis0.github.io/series/f…
1
3
78
one thing a voice PR review caught: talking over the agent's hold message could cancel the text it was about to send.
the caller was interrupting the speech. the text-send action could get cancelled too.
launched my first linkedin newsletter and 200+ people subscribed in under 12 hours!
it's for engineers finding their footing in AI: using the software skills you already have to build AI products, and using AI to help you build.
join me: linkedin.com/newsletters/eng…
can build software, but unsure where to start with ai?
i wrote 3 illustrated guides that build on each other:
1. where the model fits
2. how a request becomes a real action
3. what changes when the user speaks
the reading path starts here:
deepalis0.github.io/
pick the guide closest to what you're building: a first ai product, a tool-using assistant, or a voice agent.
start here:
deepalis0.github.io/
want the next software-to-ai lesson? subscribe to my linkedin newsletter:
linkedin.com/build-relation/…
62
the easiest way to get misleading results from an agent evaluation? test the wrong agent.
in an early learning loop demo, my simulator called a generic model instead of the agent i meant to test.
the conversations looked plausible. they came from the wrong code path.
for a concrete simulation setup, langsmith's guide separates the simulated user from the application being tested.
that boundary is worth inspecting before trusting the resulting conversations.
docs.langchain.com/langsmith…
1
1
16
how do you check that an eval ran the intended agent version?
my guides connect model behavior, tools and testing:
deepalis0.github.io/
for more software-to-ai lessons, subscribe on linkedin:
linkedin.com/build-relation/…
1
14
Hear me out: If you’re a high performer who genuinely loves building and solving complex engineering problems, have a side project!
Otherwise, corporate performance punishment can make you regret being good at your job.
Learned this the hard way!
3
176
One thing I learned while building evals for AI agents:
Fixing one bug can quietly introduce another.
You fix a tool-calling issue, but suddenly the agent’s tone changes or it starts repeating itself.
Solution? Run both qualitative and quantitative evals !
13
9
476
You need to evaluated for workflow correctness which includes tool call trajectory correctness .
1
76
built an AI chat agent, but how do you validate it before releasing to real users?
Happy-path tests won't tell you what happens when users get frustrated, stay vague, or reveal important details halfway through.
I used simulated users to test exactly these scenarios.
#evals
Being cut off by a voice agent is a frustrating user experience when the user is looking up a detail on their phone.
A pause doesn't always mean they're done.
This Voice AI Guide explores why faster replies can come at the cost of interruptions:
deepalis0.github.io/series/v…