# about.md
β
ML Researcher Β· Co-Founder Β· Mumbai β Atlanta
## Quick facts
- University: "Georgia Tech, CS '29"
- IB_Score: 43/45 // top 1% worldwide
- ICSE_CS_Rank: 1 // All India
- Location: "Mumbai, India"
- Stack: "Python Β· React Β· TypeScript Β· PyTorch Β· R Β· SQL Β· Swift"
- Research: "Game intelligence Β· HMM forecasting Β· CTC transcriptomics Β· LLM evals"
- Building: "Soverin" // see work.json
// work.json
{
"experience": [
"role": "Founding Engineer",
"org": "Soverin Labs",
"period": "Jul 2026 - Present",
"desc": "Building the intelligence layer for games."
"role": "Founder",
"org": "sbl.you",
"period": "Jun 2026 - Present",
"desc": "Training, sleep, fatigue and nutrition fused into one longitudinal model of a lifter. A deterministic pipeline β Epley e1RM tracking, acute:chronic workload ratios with monotony and strain, one-week-lagged correlations between recovery and next-session performance β computes every number; Iris, the LLM layer, is structurally barred from arithmetic and may only narrate a computed verdict, so it cannot hallucinate a lift. Expo/React Native + Supabase with row-level security, HealthKit ingest, Hevy/CSV import."
"role": "ML Researcher",
"org": "Independent (Dr. Sarita Poonia)",
"period": "Apr 2026 - Present",
"desc": "Building multi-stress transcriptomic framework using GNNs to identify pan-stress survival hub genes in CTCs and predict immune evasion from gene expression data."
"role": "Co-Founder & CTO",
"org": "IB Cognito",
"period": "Jul 2024 - Present",
"desc": "Built free IB learning platform from scratch β 10k monthly users, 108 countries. Backed by WACP International."
"role": "Co-ops Lead",
"org": "MLV",
"period": "Feb 2026 - Present",
"desc": "Managing and scaling brand ambassador network across India. Recruiting, onboarding, and performance-tracking community leads."
"role": "AI/ML Intern",
"org": "Pharmint",
"period": "Jun - Aug 2025",
"desc": "ML classification of hypoxic vs normoxic cancer cell states from RNA-seq data. 83.3% accuracy via LOO CV. Supervised by Dr. Sarita Poonia."
"role": "Web Developer",
"period": "Jun - Sep 2025",
"desc": "End-to-end dev of IB educational platform with 3k+ paid users. 40+ dynamic pages, interactive quizzes (HTML/CSS/PHP), Stripe payment infra."
"role": "Policy Research Intern",
"org": "Government of Uttar Pradesh",
"period": "May - Jul 2025",
"desc": "Proposed cooperative-based economic model for farmers in Hathras district. Focused on feasibility and long-term sustainability."
"role": "Library Development Manager",
"org": "Share A Book India",
"period": "Mar 2025 - May 2026",
"desc": "Launched city's first Library Development Program. Impacted 100+ underprivileged students, improving English reading by at least one grade."
"role": "Teaching Intern",
"org": "Schoolhouse.world + Udaan India Foundation",
"period": "Jul 2024 - Jun 2025",
"desc": "SAT Math to 40+ learners across 20+ countries (Schoolhouse). Science & reasoning to 40+ underprivileged students (Udaan)."
]
}
// projects.ts β click any block to expand
const projects = [
// currently building
Soverin Labs
β stealth
Founding Engineer Β· Jul 2026 β present
Building the intelligence layer for games.
stack: [ββββββββ, ββββββ],
details: βββββββββββββββββββ
// under wraps β more soon
βΆ
{
name: "sbl.you",
status: "in development π’"
500+waitlist
Build3accepted Β· declined
0LLM-computed numbers
// the architectural constraint
logs β rules engine + stats β verdict β Iris(narrate)
// the LLM never touches the arithmetic β it receives a computed verdict and renders prose
desc: "Most AI fitness apps hand the model your logs and hope. This one makes that impossible by construction. A deterministic pipeline owns every number: per-exercise tonnage and Epley e1RM tracking, training load as a 7-day acute against a 28-day rolling chronic with monotony and strain, RIR calibration, stall detection, and one-week-lagged Pearson correlations between recovery signals and next-session performance β each gated on a minimum of paired weeks, carrying a directional prior, and reported with a confidence tier that scales with sample size rather than a bare p-value. Only after the verdict is computed does Iris β the LLM layer β get called, and it receives the finished numbers as read-only context. It has no arithmetic path, so a hallucinated PR is not a bug class that exists here."
stack: ["Expo / React Native", "TypeScript", "Supabase + RLS", "Postgres", "HealthKit"],
ingest: ["Apple Health (sleep, nutrition)", "Hevy", "CSV"],
traction: "500+ waitlist Β· accepted into Build3, chose not to proceed",
next: "sbl.us β coach-facing tier with client dashboards"
},
βΆ
{
name: "Gaussian HMM Football Predictor",
status: "live-tested on WC 2026"
65.4%all 104 WC matches
82.1%84 decisive matches
33,455matches trained on
// the form ladder β 7 latent states, learned not hand-written
Poor β Below Avg β Weak β Average β Above Avg β Strong β Elite
// a 7Γ7 heatmap of state means over the 7 form features recovers this ordering from the data alone
desc: "Elo answers who is better. It cannot answer who is in form. This models form as a latent Markov state per team, inferred by the forward algorithm from seven observable symptoms β EWA win rate and goal difference, results against strong opposition, scoreline volatility, and both momentum terms. The full posterior over all 7 states is kept rather than collapsed to a point estimate; the 7Γ7 joint grid for a fixture, plus each side's confidence and entropy, becomes a 60-feature description that a logistic head maps to W/D/L."
interpretability: [
"Form ladder β 7Γ7 heatmap of standardized state means; the PoorβElite ordering is recovered, not imposed",
"Form trajectory β stacked-area of a team's posterior across a tournament, watching form move rung to rung; a scalar rating cannot show this",
"Reliability diagram β predicted vs observed frequency for W/D/L, overlaid on Elo; the case for a calibrated probability over a point estimate"
],
stack: ["Python", "hmmlearn", "scikit-learn", "McNemar / bootstrap CIs"],
note: "Honest limitation: on raw pick accuracy it ties Elo β they agree on 205 of 211 matches. The edge is probability quality, not picking more winners: explicit draw probabilities, calibration, and a confidence knob Elo has no equivalent of."
},
βΆ
{
name: "BroScienceBench",
status: "benchmark v1.1 Β· frozen"
246items
8models evaluated
0.89Gwet's ACβ (judge)
desc: "An LLM benchmark for gym misinformation β confident, popular, evidence-contradicting training and nutrition claims. Core finding: multiple choice saturates (7 of 8 models hit ~100%) and hides everything; the open-ended protocol opens a real 76β98% gradient. Scored for both accuracy and myth-adherence, graded by a judge outside every evaluated model family."
stack: ["Python", "OpenRouter", "Wilson CIs", "exact McNemar"],
release: "HuggingFace dataset + datasheet"
},
βΆ
{
name: "Circadian & Hypoxia Axes in CTC Metastasis",
status: "research"
306single CTCs integrated
p=0.0036rest-phase proliferation
p=0.41hypoxia Γ circadian β independent
// the result: not one coupled pathway, but two unlinked ones
circadian timing β₯ hypoxic survival
// the apparent coupling was Simpson's paradox + leukocyte contamination
desc: "Rest-phase CTCs are known to be more aggressive; tumor hypoxia and neutrophil escort are separately implicated in metastasis. Whether these form one coupled pathway or independent mechanisms was unresolved. Integrating 306 single breast-cancer CTCs across xenograft and primary patient models shows they are entirely independent axes: rest-phase CTCs are significantly more proliferative (p=0.0036, consistent across all tumor sources), and a GSEA-validated hypoxia-survival program (HALLMARK_HYPOXIA, adj. p=0.02) shows no correlation with circadian phase (p=0.41)."
key_finding: "The apparent links between hypoxia, CTC clustering and neutrophil escort are computational artifacts. Source-aware modeling shows they arise from a Simpson's paradox rooted in differing baseline hypoxia across tumor models, compounded by leukocyte-specific contamination (S100A9) β together manufacturing the illusion of one unified mechanism.",
implication: "Systemic circadian timing and local microenvironmental stress drive metastasis through distinct, unlinked programs β and single-cell oncology needs source-controlled analysis to avoid inventing mechanisms.",
stack: ["R", "Seurat", "Harmony", "fgsea", "MSigDB Hallmark"],
supervisor: "Dr. Sarita Poonia"
},
βΆ
{
name: "Hot Corners",
status: "shipped"
desc: "A macOS menu-bar app that launches any app when you flick the cursor into a screen corner. Per-corner app assignment and activation delay, no dock icon. Built because macOS only lets you bind corners to system actions, not to apps."
stack: ["Swift", "AppKit", "SwiftPM"]
},
βΆ
{
name: "IB Cognito",
status: "live π’"
10k+monthly users
108countries
WACPbacked
desc: "Free IB learning platform built from scratch. Global reach spanning 108 countries β zero marketing budget."
stack: ["React", "Node.js", "PostgreSQL"],
},
βΆ
{
name: "Black-Scholes Options Pricer",
status: "complete"
// Black-Scholes
C = SΒ·N(dβ) - KΒ·e-rTΒ·N(dβ)
// norm_cdf via Abramowitz & Stegun (err < 7.5e-8)
desc: "European option pricing from scratch. Built analytical Black-Scholes, all 5 Greeks (Ξ Ξ V Ξ Ο), Monte Carlo via geometric Brownian motion β zero financial libs."
stack: ["Python", "NumPy", "Monte Carlo"],
period: "Mar - Apr 2026"
},
βΆ
{
name: "Ensemble Stock Forecaster",
status: "published"
55%directional acc
150+downloads
5models ensembled
desc: "Multi-model ensemble integrating RF, XGBoost, ARIMA, Prophet, Gradient Boosting on SPY data. 15+ technical indicators. Published in RARS."
stack: ["Python", "XGBoost", "Streamlit", "yfinance"],
paper: "Research Archive of Rising Scholars Β· Sep 2025"
},
βΆ
{
name: "ML Cancer Genomics Classifier",
status: "research"
83.3%accuracy
9,908genes analyzed
LOOcross-validation
desc: "Random Forest on high-dimensional gene expression data to classify hypoxic vs normoxic cancer cell states. PCA + DEA on GEO GSE260794."
stack: ["Python", "scikit-learn", "PCA", "RNA-seq"]
},
βΆ
{
name: "RSA Decryption Time Modeller",
status: "IB IA Β· A grade"
// Pollard's Rho complexity
T β O(n1/4) vs O(n1/2) naive
// 308-digit key β 10^313s (naΓ―ve) | 10^152s (Rho)
desc: "Full RSA pipeline from scratch. NaΓ―ve Trial Division + Pollard's Rho factorization engines. Benchmarked across 4-13 digit primes, extrapolated to real-world 308-digit keys."
stack: ["Python", "Cryptography", "Number Theory"]
}
]
export default projects;
# awards.yml
honors:
Principal's Award
Awarded to 6 of 100 students for all-rounder excellence across academics, leadership & extracurriculars.
Young Innovator's Award
For building IB Cognito and scaling to 10k monthly users while studying IBDP full-time.
4th Place β Harvard Stock Pitch
International equity pitch competition. Delivered valuation analysis, growth catalysts, and risk assessment to Harvard panel.
AMO 2024 Bronze Β· World Rank 17
American Math Olympiad. Bronze medal, ranked 17th globally and 14th in India.
All India Rank 1 β ICSE CS
Top of India in ICSE Computer Science examinations.
IRO 2025 Honor Roll
Recognized for research excellence on the international stage.
NJCO 2024 Bronze
Bronze medal at the National Junior Chemistry Olympiad.
Published Researcher (Γ2)
SPY forecasting paper (150+ downloads) + Quantum Cryptography paper in IERJ.
#!/bin/bash
# contact.sh
$ echo "Let's build something."
β Always open to research collabs, builder chats, or just a good idea.
$ currently
β Building Soverin β the intelligence layer for games
β Researching CTC immune evasion w/ GNNs
β CS @ Georgia Tech