nandini kansal · agentic systems engineer

i build agents that do real work, and measure where they break.

Engineer at GoDaddy shipping workflow orchestration at platform scale. Formerly founding CTO of Overcast, where autonomous agents ran incident response end to end. Founder of Soma. Author of sre-bench, trajeval, and a study on how a single tool error becomes a permanent false memory. Everything below is filmed, published, or reproducible.

github x / nan.jpg linkedin email resume
scroll · the particles follow the story
01 · selected work

systems you can
watch think.

Two agent platforms and two published benchmarks. The veil lifts below. Every panel replays a real run, and the films are one click away.

01

selected work

filmed demos and reproducible numbers
01 founding cto · nov 2025 to apr 2026

overcast: production breaks, agents open the PR

An autonomous production reliability platform I built end to end as founding CTO. SDKs stream logs, traces, and metrics from inside the application. A LangGraph agent pipeline detects the incident, correlates across services, performs root cause analysis with an evidence timeline, and opens a GitHub pull request with the fix. A quick check gate held LLM cost to roughly $0.004 per automated incident.

langgraphfastapipostgres + redisreact + tsmulti tenant auth
$0.004
cost per incident
6
agents in pipeline
<2 min
detect to PR
incident pipeline · simulation
checkout service
a compressed replay of the real pipeline watch the film ↗
02 founder and sole engineer · usesoma.app

soma: an agentic personal OS

A multi step agent loop on the Claude API with tool use, model routing, and per run budget caps. Long term memory through contextual retrieval. A nightly learning agent that updates its model of you while you sleep, graded by a trajectory eval harness that scores tool choice, arguments, and end state separately. And a physical plant that thrives or droops with your alignment score, because software that knows you should be able to look you in the eye.

live productclaude apicontextual retrievalreact nativesupabaseraspberry pi
4
agents in the loop
nightly
learning runs
1
living plant
usesoma.app →
agent trace · nightly brain run
soma
every soma run produces a trace like this watch the film ↗
03 open source · sre-bench + trajeval

sre-bench: break it on purpose, grade the world

Four dockerized microservices, a chaos injector that induces realistic outages, and six scenarios with traps that separate diagnosis from flailing. Agents are scored by trajeval, my own trajectory eval library, on tool selection, argument correctness, and end state, verified by probing the environment rather than trusting the agent's own claims. Every failure on the board is a confident agent contradicted by reality.

dockerchaos injectiontrajectory evalsclaude agent sdkreproducible
6 × 3
scenarios × seeds
3
models on the board
0.94
top overall score
sre-bench repo → trajeval →
chaos run · scored trajectory
runaway-retry · seed 1
the probe grades the world, not the claim watch the film ↗
04 research · paper + reproducible repo

transient errors, durable beliefs

Does one ordinary tool error become a permanent false memory in an agent's consolidated long term memory? Yes, and the mechanism sits in the pipeline's timing rather than the model's stubbornness. A single stale tool return is committed as a durable belief, retained without decay, and followed by every dependent decision downstream, while the agent re-verifies in 1.3% of 1,427 opportunities. Confronted directly with contradicting evidence, agents update every time and defend nothing: this is a memory architecture failure, not a personality flaw. One enforcement rule, that flagged facts must be re-verified before use, bounds a false belief's lifetime to one or two cycles.

agent memoryconsolidationbootstrap CIstranscript auditedMIT licensed
1.3%
re-verification rate
70/70
update, zero defense
≈$15
total study cost
read the paper (pdf) → repo →
commit rate vs consolidation interval
n=10 per cell
faster consolidation, more vulnerable
02 · production

an ocean of traffic.
I've been paged for it.

GoDaddy's Airo platform, 2024 to present. The waves below are calmer than the incidents were.

02

production scar tissue

godaddy · airo platform · 2024 to present
long running orchestration

cold hibernation, shipped end to end

A two tier archive and restore system for inactive apps on AWS Step Functions: backup validation gates, cross cell restores with secret migration, distributed locking. Cut per app infrastructure cost for the archived fleet by 75%. Owned across four services from design through rollout.

incident response

the 25,000 event ceiling

A DNS verification retry loop quietly hit the Step Functions execution history limit in production. I diagnosed it and redesigned the retry semantics. Unbounded loops are the agent problem too, which is why every agent I build ships with a hard step cap and a budget gate.

reliability engineering

races, leaks, and coverage

Root caused a backup registration race behind intermittent restore failures. Shipped an orphaned resource cleanup pipeline nobody had ticketed. Drove test coverage from 70 to 90 percent with enforcement that stuck.

agents at work

runbook agent prototype

Hackathon build with the network operations team: an agent that reads live incident context and proposes remediation grounded in runbooks. The seed of Overcast's remediation design, and of every scenario in sre-bench above.

03 · shipping now

one build in flight,
three students racing.

Can a quantum solver pick a better lesson plan than a greedy one? The curves publish either way.

3
solvers, one teacher
weekly
LoRA cycles
03

shipping now

built in public · summer 2026
in progress

qurriculum: quantum curriculum optimization

A frontier agent whose job is raising a small open model. Each week it evals its student, builds a weakness profile, generates a candidate pool of training examples, and then has to choose the batch. That choice is a combinatorial problem: maximize coverage of what the student got wrong, maximize diversity, penalize redundant pairs, respect a token budget. Binary include or exclude decisions with pairwise penalties is a QUBO, so the batch gets selected three ways in parallel, by a greedy baseline, a classical annealer, and a quantum annealer, and three students train on the three curricula. The existing literature scores the solver's own objective. This scores downstream student learning, closed loop, where this week's eval reweights next week's problem.

follow the build →
preview the report card

an agent, raising a model

Every week the teacher fine tunes all three students, runs the evals, and files a report card. Quantum losing is a publishable result, which is the point: the comparison is the contribution, not the outcome. The format below shows sample data until the first semester's curves land.

first real curves: fall 2026
report card · sample format
student score
sample data · not results

writing

building intelligent things without losing yourself
transient errors, durable beliefs: false memory in agent pipelinespaper · pdf
the frontier model doesn't top the board: sre-bench findingsleaderboard · july 2026
grade the world, not the model's claimstrajeval · docs
everything else, live on x →nan.jpg