@umdcs research @ gel
Joined December 2017
- Tweets 11,093
- Following 1,144
- Followers 1,173
- Likes 20,738
959 Photos and videos
Pinned Tweet
Jun 11
we formalized this further in long horizon research tasks, specifically where there are multiple goodness of fit metrics that need to be considered and optimized for.
one thing we realized is that the loop is most valuable at the accept/reject boundary.
a global aggregate score can improve while the result moves in the wrong local direction, so the external loop has decide if the agent did actually find a better solution, or did it just find a local tradeoff that makes the headline metric look better?
the paper explore's this on a mechanistic ecology model used in NASA’s Carbon Monitoring System.
arxiv.org/abs/2606.11522
1
1
10
2,988
Jun 11
we formalized this further in long horizon research tasks, specifically where there are multiple goodness of fit metrics that need to be considered and optimized for.
one thing we realized is that the loop is most valuable at the accept/reject boundary.
a global aggregate score can improve while the result moves in the wrong local direction, so the external loop has decide if the agent did actually find a better solution, or did it just find a local tradeoff that makes the headline metric look better?
the paper explore's this on a mechanistic ecology model used in NASA’s Carbon Monitoring System.
arxiv.org/abs/2606.11522
1
1
10
2,988
Jun 11
refer this tweet thread for more information about the ecology model and autoresearch's impacts.
x.com/devparagiri/status/205…
May 11
for the past few weeks, i spent some time trying to formalize this approach further and get it adopted as a real tool with my lab.
the main extension lies in functional form search hyperparam tuning in a tractable manner with ablations to prune overt complexity.
we specifically focused on improving our model's ability to better explain fire and our offline results now top the benchmark!
a more detailed blog, focusing on the approach transferable learnings can be found here: paragiri.com/blog/2026/autor…
137
Jun 7
i need an arxiv endorser for cs dot AI. dm me if u can or know someone!
3
150
May 13
interesting concept
May 13
Today, we’re launching pgGraph: an Apache open-source graph traversal engine for Postgres, written in Rust 🦀
When we started Evokoa, we kept running into the same wall every serious agent team eventually hits:
> Agents need to reason across relationships.
> But graph DBs are expensive AF, and suck to use.
So, we built pgGraph around a simple, single idea:
Postgres should stay the source of truth, and the graph engine should live beside it.
pgGraph does something different.
> It keeps the rows in Postgres,
> Compiles the topology around them,
> Creates a virtual graph layer using CSR-style adjacency arrays.
We're making your existing postgres database graph-traversable for agents without any of the usual BS.
> No recursive join hell.
> No ETL pipeline.
> No second source of truth.
> Blazing fast performance
pgGraph is already live in production workflows across RevOps, healthcare, and visa services.
We’re open-sourcing it because graph traversal should become a default primitive in the agent stack, not an enterprise migration project.
Treat Postgres as a graph.
Zero data migration.
This is what Apache AGE should have been.
Docs Repo below.
1
5
252
May 13
how are they letting everyone be on the same damn plane?
BREAKING: President Trump says "the great" Jensen Huang of Nvidia, $NVDA, is currently on the Air Force One with him on the way to China.
Trump says Elon Musk, Tim Cook, Larry Fink, Stephen Schwarzman, David Solomon, and many other CEOs are joining him on the trip.
4
220
May 12
endgame
May 12
NEW: Meta employees in the U.S. are organizing a protest over mouse-tracking technology.
1
1,200
May 12
this is very cool, exciting times!
May 11
Today we're sharing our work on interaction models. A new class of model trained from scratch to handle real-time interaction natively, instead of gluing it onto a turn-based one.
youtu.be/A12AVongNN4
1
235
May 11
for the past few weeks, i spent some time trying to formalize this approach further and get it adopted as a real tool with my lab.
the main extension lies in functional form search hyperparam tuning in a tractable manner with ablations to prune overt complexity.
we specifically focused on improving our model's ability to better explain fire and our offline results now top the benchmark!
a more detailed blog, focusing on the approach transferable learnings can be found here: paragiri.com/blog/2026/autor…
Mar 20
i extended this paradigm to earth system models. ed v3.0 simulates plant growth, fire, soil carbon, and water cycling globally using parameterized formulas, many unchanged since the 90s. ilamb benchmarks it against 21 other models.
for each submodel, the system searches over both formula structure and continuous parameters via bayesian optimization. every candidate equation must map to a named physical mechanism.
the system also selects the appropriate goodness-of-fit metric set per module. optimization runs in phases following the model's dependency graph since upstream modules (photosynthesis) feed into downstream ones (soil carbon, fire).
results (spatial correlation r = how well the model's predicted global map matches gridded observations):
to read the detailed implementation and improvements, blog link is attached below!
1
12
3,661
May 5
got to nyc this saturday evening.
met @jonathanzliu, @itsmnjn and @anaclumos early sunday. my entire week has been mayhem since then.
have like 10 meetings with a booked schedule until coming saturday.
i love @x and this city. excited to move here soon!
1
4
200