$ init utkarsh.saraogi --target=portfolio
[ ok ] mounting /skills 82 chips loaded
[ ok ] connecting github://utkarsh1610-s 6 repos
[ ok ] bootstrapping lineage.dag 4 layers, 31 nodes
[ ok ] warming stream.console subscribed → 6 topics
[ ok ] ready. press / to query.
LIVE
/
utkarsh.saraogi
/
Boston, MA
/
—
open to co-op · Spring '27
query /
open to Spring/Summer '27 co-op
Utkarsh Saraogi

Utkarsh
Saraogi.

$ role =

Three years at HPE taught me that data systems don't fail because of bad ideas — they fail because of bad data. That experience shaped how I think as a data engineer. I have worked on the reliability side of a production data platform — tracing where data broke and why. Now I aim at building data systems end to end.

As an alternate interest area, I have done substantial research on why LLMs quietly fold when you push back on them. Now, aiming EACL 2026 submission.

view work → / query portfolio github ↗
stream://utkarsh.events
consuming
0 events
topic: portfolio.v3
lag: 0ms
offset 0
01

snapshot.

what the numbers say
facts ingested 181M SEC XBRL · 69 quarters · bitemporal warehouse
look-ahead bias 3.3% top-decile picks that differ · 49 quarters
records processed 46M+ ELT · 3 raw sources · 70+ countries
recovered $7.8M counterfactual A/B · credit risk · 113K loans
MTTR reduction −40% 15+ microservices · telemetry triage
streaming latency <60s MarketPulse · Kafka → Spark → BigQuery
02

lineage.

how raw experience flows to gold
raw BRONZE
▢ HPE · 3 yr cloud dev 2022–25
▢ Northeastern MSCS 2025–27
▢ NLP 2026
skills SILVER
⊡ Jenkins CI/CD
⊡ Docker · containers
⊡ CrowdStrike telemetry
⊡ DevOps · GCP · AWS
⊡ Kafka · Spark Streaming
⊡ PySpark · dbt · Airflow
⊡ Databricks · Lakeflow
⊡ Hidden-state probing
⊡ Causal · A/B · SHAP
projects SILVER
▣ Hindsight
▣ MarketPulse
▣ SoundStream Analytics
▣ Credit Risk Engine
▣ Sycophancy Detection
▣ Stock Portfolio Mgr
outcomes GOLD
★ 21.8M PIT intervals
★ −40% MTTR
★ shipped to prod
★ 4K+ failure labels
★ ACL ARR submission
★ $7.8M recovered
★ sub-60s latency
tip hover any node to trace its upstream + downstream dependencies →
03

projects.

deployable, instrumented, in production
one fact, two truths AT&T · FY2021 revenue · as the market knew it restated 2023-02-13 FY2022 10-K · WarnerMedia spin-off $168.9B $134.0B as_of('2022-06-01') as_of(today) 2022 2024 2026 21.8M knowledge intervals · 14,885 companies · 6 invariants asserted
Data Engineering shipped

Hindsight — a warehouse that remembers what people knew, and when

Companies quietly correct their past financial results, and every database overwrites the original. Backtest on that and you're trading on numbers nobody had at the time. Hindsight ingests 69 quarters of SEC XBRL filings (181M facts) through a Lakeflow medallion pipeline, then resolves per-fact filing precedence into 21.8M knowledge intervals — so any historical date returns only what was publicly knowable then. Measured the resulting look-ahead bias at 3.3% of top-decile stock selections across 49 quarters.

181Mfacts
21.8Mintervals
0.001%broken rows
databrickslakeflowdelta lake auto loaderunity catalogpysparkpytest
→ source → read the writeup
KAFKA trade.events ticker-keyed SPARK structured streaming VWAP 1min · 30s wm 3 detection signals bronze raw trades silver vwap windows gold anomalies BQ dbt p99 ≈ 58ms · 5 instruments · sub-60s end-to-end 3 detection signals · 8 dbt tests · fault-tolerant checkpointing
Data Engineering live

MarketPulse — real-time market anomaly pipeline

Equity trade events stream through Aiven Kafka (SSL/TLS, ticker-keyed) into PySpark Structured Streaming with 1-min tumbling VWAP windows and 30s watermarking. Medallion lakehouse on GCS with 3 detection signals — Z-score volume spike, intra-window price deviation, wash trading — each with per-signal confidence. Gold lands in BigQuery via 3 dbt models with 8 passing data quality tests.

<60slatency
5instruments
8/8dbt tests
kafkaspark structured streamingdelta lake gcsbigquerydbtdocker
→ source
NLP · Research ACL ARR · EACL 2026

Sycophancy under pressure — predicting LLM capitulation

First-authored paper investigating hidden-state representation shifts in LLMs before sycophantic flips. Novel monotonically escalating pressure schedule across 5 architectures and 150+ probes. Cosine similarity disruption at the first pressure turn predicts behavioural flip 3–4 turns ahead — without any probe training.

L28dominant layer
75.5%probe acc.
+11.7over chance
pytorchqwen2.5-7b hidden state probingrbf svm
↓ research deep-dive → github
Data Engineering live demo

SoundStream Analytics — 46M record music ELT

Batch ELT pipeline processing 46M+ records from Spotify Charts, Spotify Songs and Last.fm. Medallion architecture orchestrated as an Airflow DAG; 60% storage compression via partitioned parquet; Silver-layer quality checks dropped 5,043 duplicates across a 36.7% cross-source match rate.

46M+records
60%compressed
70+countries
pysparkdbtairflow duckdbstar schemastreamlit
→ source → live demo ↗
Data Science

Credit Risk Strategy Engine

Hypothesis-driven EDA on 113K+ loans. Kaplan-Meier survival revealed 70% of defaults show distress within 12 months. HistGradientBoosting on 14 bureau features + counterfactual A/B framework recovered $7.8M in loan volume with 20% default reduction. FCRA-compliant SHAP adverse-action codes.

113Kloans
$7.8Mrecovered
-20%defaults
histgradboostshap counterfactual a/bkaplan-meier
→ source
Data Engineering

Stock Portfolio Manager

End-to-end pipeline pulling real-time market data from Finnhub API → transformed via SQL stored procedures → normalised MySQL → served as portfolio analytics through a Flask REST API with a React frontend.

RESTapi
MySQLstore
Reactui
pythonmysql flaskreactfinnhub
→ source
Build you're here

this site — a queryable portfolio

The site you're reading is itself a small experiment: a portfolio that behaves like one of my pipelines. Stream console replaying real visitor events, interactive lineage DAG, live interpretability widget, and a SQL-ish command palette that actually queries the page. Press / to try.

htmlcssvanilla jsno framework
→ open console
04

research.

when the model caves — and how to see it coming
ACL ARR · May '26 Northeastern

"Before the Model Caves —
Detecting Pre-Capitulation States in Multi-Turn Sycophancy"

Sycophancy in LLMs is usually studied after the model flips. We asked the opposite question: can you tell, from the hidden states alone, that a flip is about to happen — before any visible change in the output?

We built a monotonically escalating pressure schedule across five turns and traced hidden-state activations through every transformer layer. The signal was there, and it was loud. In Qwen2.5-7B, Layer 28 dominates; across all five models, layers 17–19 are globally predictive — a dissociation unreported in prior work.

key finding cosine similarity disruption at the first pressure turn predicts behavioural flip 3–4 turns in advance, with no probe training required. RBF SVM + KNN classifiers reach 75.5% accuracy — +11.7 over chance.

First author with Soham Padia (equal contribution); with Tomas D'Avola, Vedant Shah; advised by Prof. Malihe Alikhani. Targeting EACL 2026.

qwen2.5-7b · layer activations L2 norm, normalised
pressure schedule ← drag through turns to see hidden state →
activation pattern at turn 0
stable · model holding original position
05

experience.

log // tail -f career.events
Sep 2025 — present
Boston, MA
Graduate Researcher Sep 2025 — present
NLP interpretability group. Multi-turn sycophancy — why language models abandon a correct position under conversational pressure, and whether the collapse is visible in hidden states before it reaches the output.
5architectures
150+probes
75.5%probe acc.
3–4turn lead
  • First-authored ACL ARR submission on detecting pre-capitulation states in multi-turn LLM sycophancy — targeting EACL 2026
  • Engineered the monotonically escalating pressure schedule that became the paper's central novelty; ran the experiment matrix across 5 architectures × 150+ probes
  • Localised the predictive signal: Layer 28 dominant in Qwen2.5-7B, layers 17–19 globally predictive — a dissociation unreported in prior sycophancy literature
pytorchqwen2.5-7b transformersrbf svmnumpyhidden-state probing
Jan 2022 — Aug 2025
Bangalore, IN
Cloud Developer - Hewlett Packard Enterprise Jan 2022 — Aug 2025 · 3 yr 8 mo
Switch Data Monitoring — the telemetry platform ingesting hardware health and network metrics from HPE's enterprise switch fleet across 10+ models, through an enricher → predictor → sink connector chain. I worked the reliability side: where data broke, why, and keeping validation honest as the platform moved underneath it.
15+microservices
−40%mttr
−55%cycle time
4K+failure labels
Cloud Developer Aug 2022 — Aug 2025 · 3 yr 1 mo
  • Traced data loss across 15+ microservices ingesting VLAN/VXLAN switch telemetry, following records through enricher → predictor → sink connectors to pinpoint the failing service and root cause → −40% MTTR
  • Built training dataset from 4,000+ failure logs, designed annotation schema for an ML triage assistant → +15% classification accuracy, removed manual escalations
  • Designed a parallel execution framework distributing validation scripts across heterogeneous switch fleets → −55% end-to-end cycle time
  • Owned Jenkins CI for the platform's test suite and lab environments; kept validation coverage current as GraphQL APIs and deployed services drifted from test expectations
  • Mentored 3 interns on automation workflows; documentation reduced onboarding by 1–2 months
Software Engineer Intern Jan 2022 — Jul 2022 · 7 mo
  • Built automated feature-validation scripts integrated into the CI pipeline across 10+ switch models; eliminated recurring production rejections from malformed telemetry payloads
  • Cross-referenced 500+ historical debugging logs to verify log-extraction accuracy; authored the Confluence reference adopted for stack onboarding
pythonjenkins ci sqlgraphqlcrowdstrikedockerconfluence
06

education & off-screen.

how I got here and what I do when the laptop is closed
Northeastern University
MS · Computer Science · Boston, MA
Sep 2025 — May 2027 (expected)
coursework MLOps · NLP · Data Mining · Database Management Systems · Algorithms for Data Science · Deep Learning
Manipal Institute of Technology
BTech · Electronics & Communication · India
May 2022 minor — data science & modelling
coursework Data Structures · Data Science · Python · Cloud Computing · Data Modelling
[01]

football — vice-captain, weekend referee

Vice-captain of MIT Manipal's inter-college team (20+ players, regional tournaments). Currently refereeing competitive matches on weekends — 10+ years of playing. The pattern-recognition transfers cleanly to production triage.

[02]

HPE promotion track

Approved for promotion based on contributions to the automation framework and the ML-powered triage initiative. Multiple team recognition badges.

[03]

student governance

Member of the Official Student Committee at MIT Manipal — campus initiatives, event ops, liaison between students and administration.

[04]

off-keyboard

Football. Movies and shows. Tech blogs. Pattern-matching across domains is the actual job.

let's build something.

Open to Spring '27 and Summer '27 co-op + internship roles
in Data Engineering, Data Science, and ML Engineering.

→ email me / open query console
email saraogi.u@northeastern.edu linkedin /in/us1610 ↗ github /utkarsh1610-s ↗