calibrating sensors …

AI/ML Engineer · Applied AI

I find signals
in noise.

AI/ML engineer. I build systems that pull faint truths out of overwhelming data — from operating rooms to distant stars.

SCENE 01
CHAPTER 01 — THE SKY

A planet's shadow in 144,000 light curves

NASA's TESS telescope stares at the sky and streams back noise — stellar flicker, instrument drift, cosmic junk. Somewhere in it, a star dims by a fraction of a percent: a planet passing in front. COSMOS hunts that dip.

Method

An end-to-end detection pipeline on consumer hardware: 15 orthogonal detectors — classical Box-Least-Squares transit search, the MOMENT time-series foundation model, and dedicated channels for single transits, exocomets and microlensing — swept across the TESS survey. Honest verdict included: candidates reported as candidates, nulls as nulls.

144k
TESS light curves surveyed
15
orthogonal detectors
63.9%
known-planet recovery (ensemble)
PythonBLSMOMENT / ChronosGPU pipeline
SCENE 02
CHAPTER 02 — THE BODY

The crash, 4½ minutes before it happens

In an operating room, vital signs whisper before they scream. PulseGuard listens to the whisper — an early-warning system that flags intraoperative deterioration minutes before it becomes a crisis, without drowning clinicians in false alarms.

Method

Trained on VitalDB (5,170 real anesthesia cases, Seoul National University Hospital), externally validated on MOVER (UC Irvine) — two hospitals, two continents. Gradient-boosted models over routine monitor channels with conformal calibration; evaluated on a locked held-out test of 303,909 windows. Plus a live monitor streaming a FHIR feed with a calibrated 10-minute forecast cone.

~4.5 min
warning lead time
308×
fewer false alarms
0.849
AUROC, locked test
2
hospitals, cross-validated
Live monitor Code
PythonXGBoostConformal predictionVitalDB · MOVERFHIR
SCENE 03
CHAPTER 03 — THE LITERATURE

23,954 papers. The gaps glow.

Science has white space — pairs of ideas that are obviously related, that almost nobody studies together. You can't see it reading one paper at a time. CosmoScope maps an entire field from above and makes the gaps visible.

Method

Four years of arXiv cosmology and instrumentation papers embedded with sentence-transformers, clustered with BERTopic into ~39 recognizable topics. A white-space engine compares semantic proximity against actual co-citation — where proximity far exceeds co-occurrence, that's a research gap, proposed with real papers cited on each side.

23,954
arXiv papers mapped
39
discovered topics
8.6×
PTA attention surge, detected
Live dashboard Code
Pythonsentence-transformersBERTopicUMAP · HDBSCANPlotly
SCENE 04
CHAPTER 04 — THE MIND

Machines that check their own work

A language model alone will confidently invent a number. So I build systems where models police each other — one plans, one works, one judges — and none is trusted alone.

Method

Cash Flow Runway Advisor: a Planner→Executor→Judge loop with hard guardrails (iteration, token, wall-clock caps) and full trace telemetry, streaming over FastAPI. Small Action Model: Qwen2.5-3B fine-tuned with QLoRA for tool-calling, evaluated BFCL-style on decision accuracy and hallucination rate.

3
roles: plan · execute · judge
3B
params, fine-tuned tool-caller
Cashflow advisor
Claude (Sonnet + Haiku)FastAPIQLoRA / PEFTPyTorch
SCENE 05
CHAPTER 05 — THE PATH

From a satellite dish to a research lab

2022

ISRO — Indian Space Research Organisation

Research internship: satellite Earth-observation pipelines over the Bay of Bengal & Arabian Sea.

2025

Springer Nature

Co-authored book chapter: Remote Sensing Observation of Sea-Surface Parameters — born from the ISRO work.

2019–24

Shubhlaxmi Enterprises

Scaled a family distribution business 150% with Python automation; cut 10 hrs/week of manual reporting.

2026

Johns Hopkins University

MS in Business Analytics & AI — building the research lab you're scrolling through.

SCENE 06
CHAPTER 06 — PLAY & CLASSIFIED

Serious engineers also build toys

// classified — available on request

COSMOS-SBI

Dark-energy & structure-growth engine: neural simulation-based inference (SNPE + Masked Autoregressive Flows) on real survey data — ACT DR6, KiDS, DESI, Pantheon+.

PyTorch · JAX

Small Action Model

Qwen2.5-3B tool-calling LLM, QLoRA fine-tune, custom MCP tool interface, BFCL-style eval.

PyTorch · PEFT

SignalGap

Do time-series foundation models fail on irregular astronomical data? Benchmark + GP adapter: TimesFM+GP hits MASE 0.52–0.75, ~7× over raw.

Chronos · Moirai · TimesFM · celerite2

CHAPTER 07 — THE SIGNAL

The next signal
could be yours

Research collaboration, an interesting problem, or a role where faint signals matter — my inbox is open.