AYAAN
People-First Development.
Impact-Oriented Innovation.
Building Towards a Better Tomorrow.
Transforming complex technical abstractions into seamless, people-first solutions. I leverage modern machine learning and robust full-stack architectures to engineer intuitive experiences, ensuring that every intelligent system I build is defined by how well it serves the user.
Forward-pass activations propagating through a custom deep network architecture
Semantic similarity mapped as geometry in high-dimensional vector space
Distributed system blueprint — rate limiters, hashing rings, and leader election
DFA/NFA playground for stepping through automata and regular expressions
Blessed to be able to surround myself with ambitious and kind people
Statistically, I'm apart of the 99th percentile of LoL players!
Lead Software Engineer
- —Architected a locally-running Rust/Tokio HTTPS interception proxy with custom TLS termination via a device-bound CA, routing 100% of outbound LLM traffic through a detection pipeline without modifying provider authentication or breaking streaming SSE protocols.
- —Engineered a two-layer sensitive data detection system combining regex/NER pattern matching with a locally-hosted quantized Phi-3-mini model (Q4_K_M, <500MB RAM), achieving sub-150ms end-to-end sanitization latency on CPU-only hardware across OpenAI, Anthropic, and Google Gemini endpoints.
- —Built an encrypted local vault using SQLite/SQLCipher with OS keychain key derivation to store sensitive-to-placeholder mappings, enabling full response reconstruction without any client data crossing the network boundary.
- —Designed a tamper-evident audit logging system capturing per-prompt detection metadata (categories, confidence scores, SHA-256 prompt hashes) with 0 sensitive data retained, producing compliance-exportable records of all AI-assisted interactions.
Frontend Software Engineer
- —Designed and shipped tasteful, visually intensive web experiences featuring WebGL shaders, physics simulations, and scroll-driven animations, achieving Lighthouse scores >90 across audited categories, validated through Chrome DevTools performance audits.
- —Instrumented a custom Three.js WebGL renderer with Chrome DevTools, identifying a requestAnimationFrame loop consuming 722ms of GPU self-time; reduced tick self-time by 65% (722ms → 255ms) through scroll-progress-based render gating.
- —Eliminated a CLS score of 1.49, nearly 15× over Google's "Good" threshold, by migrating scroll-driven panels from layout-property animations to compositor-only transforms, achieving 100% CLS reduction site-wide with zero INP regression.
- —Reduced image payload by 89% on the primary background asset (6.7MB → 716KB) via JPEG-to-WebP conversion and re-enabling Next.js image optimization, cutting estimated per-session bandwidth by over 90%.
Learning Assistant — Calculus II
- —Support student understanding of multivariable calculus, sequences and series, and integral techniques by leading collaborative problem-solving sessions and translating abstract mathematical concepts into accessible, intuitive frameworks.
- —Bridge the gap between lecture and comprehension by identifying recurring points of confusion across students and developing targeted explanations that address conceptual gaps rather than surface-level procedure.
- —Collaborate with course instructors to align supplementary support with curriculum pacing, ensuring students build foundational fluency before advancing to more complex applications.
Executive Board Member
- —Co-coordinated a hackathon featuring $4,000+ in prizes and meaningful industry participation, managing logistics, partner communication, and attendee experience.
- —Organized a career networking event connecting 70 students with 10 professionals, navigating outreach and positioning in a campus environment dense with competing events.
- —Partnered with student leaders and external professionals to expand access to mentorship and career opportunities in tech.
Officer & Fundraiser Lead
- —Transformed a fundraiser rejection into approval by mapping admin concerns into constraints, reframing the pitch, and securing sign-off — ultimately engaging 100+ donors and raising $1,000 for humanitarian aid.
- —Aligned stakeholders across the board and school administration by translating feedback into a compliance-ready plan and updating materials to meet policy requirements.
- —Developed an execution playbook covering roles, cash-handling, and communications, enabling smooth event delivery with zero incidents.
Varsity Team Captain
- —Led a cross-skill-gap roster as the team's primary shot-caller, coordinating structured home practice sessions to develop team cohesion, communication, and strategic execution across players ranging from Bronze to Diamond-level competition.
- —Translated high-level game sense into digestible, actionable callouts for teammates several skill tiers below, building a shared strategic vocabulary that elevated collective performance beyond individual rank.
- —Guided the team to a 3rd place finish in the Garden State Esports League of Legends Championship (Fall 2022), competing at the varsity level against schools across New Jersey.
Lead Software Engineer
- —Architected a locally-running Rust/Tokio HTTPS interception proxy with custom TLS termination via a device-bound CA, routing 100% of outbound LLM traffic through a detection pipeline without modifying provider authentication or breaking streaming SSE protocols.
- —Engineered a two-layer sensitive data detection system combining regex/NER pattern matching with a locally-hosted quantized Phi-3-mini model (Q4_K_M, <500MB RAM), achieving sub-150ms end-to-end sanitization latency on CPU-only hardware across OpenAI, Anthropic, and Google Gemini endpoints.
- —Built an encrypted local vault using SQLite/SQLCipher with OS keychain key derivation to store sensitive-to-placeholder mappings, enabling full response reconstruction without any client data crossing the network boundary.
- —Designed a tamper-evident audit logging system capturing per-prompt detection metadata (categories, confidence scores, SHA-256 prompt hashes) with 0 sensitive data retained, producing compliance-exportable records of all AI-assisted interactions.
Frontend Software Engineer
- —Designed and shipped tasteful, visually intensive web experiences featuring WebGL shaders, physics simulations, and scroll-driven animations, achieving Lighthouse scores >90 across audited categories, validated through Chrome DevTools performance audits.
- —Instrumented a custom Three.js WebGL renderer with Chrome DevTools, identifying a requestAnimationFrame loop consuming 722ms of GPU self-time; reduced tick self-time by 65% (722ms → 255ms) through scroll-progress-based render gating.
- —Eliminated a CLS score of 1.49, nearly 15× over Google's "Good" threshold, by migrating scroll-driven panels from layout-property animations to compositor-only transforms, achieving 100% CLS reduction site-wide with zero INP regression.
- —Reduced image payload by 89% on the primary background asset (6.7MB → 716KB) via JPEG-to-WebP conversion and re-enabling Next.js image optimization, cutting estimated per-session bandwidth by over 90%.
Learning Assistant — Calculus II
- —Support student understanding of multivariable calculus, sequences and series, and integral techniques by leading collaborative problem-solving sessions and translating abstract mathematical concepts into accessible, intuitive frameworks.
- —Bridge the gap between lecture and comprehension by identifying recurring points of confusion across students and developing targeted explanations that address conceptual gaps rather than surface-level procedure.
- —Collaborate with course instructors to align supplementary support with curriculum pacing, ensuring students build foundational fluency before advancing to more complex applications.
Executive Board Member
- —Co-coordinated a hackathon featuring $4,000+ in prizes and meaningful industry participation, managing logistics, partner communication, and attendee experience.
- —Organized a career networking event connecting 70 students with 10 professionals, navigating outreach and positioning in a campus environment dense with competing events.
- —Partnered with student leaders and external professionals to expand access to mentorship and career opportunities in tech.
PokerMind AI — Multi-Agent Decision System (In Development)
A behavioral AI research system built around a genuinely hard problem: what does it take for an autonomous agent to play poker indistinguishably from a human? Deterministic strategy is table stakes — the deeper challenge is defeating bot-detection heuristics that fingerprint agents on action-timing regularity, bet-sizing quantization, and decision-tree periodicity. PokerMind uses a council-of-agents architecture: four LLM microservices (Strategist, Historian, Humanizer, Orchestrator) debate each action before it's committed. A behavioral evasion subsystem samples timing jitter, deliberate errors, and fatigue patterns from learned human priors. Deployed on AWS ECS with Redis message brokering and a Grafana observability layer.
Fantasy Stocks
The stock market is a lot less intimidating when it's a game. Fantasy Stocks lets you draft real companies like sports picks, join private leagues, and compete on actual market performance — built for the curious, not-yet-committed investor who wants a reason to care about tickers. Real-time drafts, weekly head-to-head matchups, and a live leaderboard make the learning feel like winning.
Sage — AI Group Chat Agent
Most group-chat bots are noise. Sage is an AI that lives inside your iMessage group and speaks fewer than 1 in 10 messages — only when it has something real to say. It answers @mentions like a friend who was there, surfaces unresolved tension when the group goes quiet, and intervenes when someone contradicts or forgets a detail established earlier in the conversation. That last behavior hints at something bigger: in large organizations where teams move fast and context gets lost, an agent with real memory could quietly prevent the kind of drift that causes costly misalignment. Built for HackPrinceton's Agents in iMessage track.
Skindex — Skin Lesion Classifier
Googling your symptoms is stressful. A clinic visit for a minor skin concern feels like too much. Skindex sits in between — snap a photo, and it classifies the condition across 10 categories with urgency levels that tell you whether to monitor it, treat it at home, or actually see a doctor. An onboard assistant handles follow-up questions. Informational support, not a diagnosis.
PokerMind AI — Multi-Agent Decision System (In Development)
A behavioral AI research system built around a genuinely hard problem: what does it take for an autonomous agent to play poker indistinguishably from a human? Deterministic strategy is table stakes — the deeper challenge is defeating bot-detection heuristics that fingerprint agents on action-timing regularity, bet-sizing quantization, and decision-tree periodicity. PokerMind uses a council-of-agents architecture: four LLM microservices (Strategist, Historian, Humanizer, Orchestrator) debate each action before it's committed. A behavioral evasion subsystem samples timing jitter, deliberate errors, and fatigue patterns from learned human priors. Deployed on AWS ECS with Redis message brokering and a Grafana observability layer.
Fantasy Stocks
The stock market is a lot less intimidating when it's a game. Fantasy Stocks lets you draft real companies like sports picks, join private leagues, and compete on actual market performance — built for the curious, not-yet-committed investor who wants a reason to care about tickers. Real-time drafts, weekly head-to-head matchups, and a live leaderboard make the learning feel like winning.
Sage — AI Group Chat Agent
Most group-chat bots are noise. Sage is an AI that lives inside your iMessage group and speaks fewer than 1 in 10 messages — only when it has something real to say. It answers @mentions like a friend who was there, surfaces unresolved tension when the group goes quiet, and intervenes when someone contradicts or forgets a detail established earlier in the conversation. That last behavior hints at something bigger: in large organizations where teams move fast and context gets lost, an agent with real memory could quietly prevent the kind of drift that causes costly misalignment. Built for HackPrinceton's Agents in iMessage track.
Skindex — Skin Lesion Classifier
Googling your symptoms is stressful. A clinic visit for a minor skin concern feels like too much. Skindex sits in between — snap a photo, and it classifies the condition across 10 categories with urgency levels that tell you whether to monitor it, treat it at home, or actually see a doctor. An onboard assistant handles follow-up questions. Informational support, not a diagnosis.
Poli — Bias Analysis Platform
A conversational platform for analyzing bias in any casual written work — news articles, blogs, opinion pieces, social media posts, political speeches. Built on the principle that truly neutral text requires no adjectives, since descriptive language is the primary vehicle through which bias enters writing. Fixates on adjective usage alongside broader linguistic patterns to score content across three dimensions: extremity, subjectivity, and factual accuracy, each rated out of 10. Motivated by a personal experiment confirming how algorithm-driven feeds create ideological echo chambers. FastAPI backend with Pydantic validation and GPT-3.5 for structured JSON analysis, a BeautifulSoup scraping layer for direct URL ingestion, and a React frontend with a chat-based interface and SVG rating visualizations.
Car Deal Predictor — Used Car Classification Model
Built a classification model in R to rate used-car deals across 10,000 listings from Dubai, achieving 92% cross-validation accuracy with XGBoost. Engineered features capturing hidden relationships between price, mileage, and benchmark value, reducing multicollinearity and improving validation accuracy by 4%. Closed a 4–5% overfitting gap through 5-fold cross-validation, model stacking, and probability-based ensembling to ensure performance held on unseen data.
5Q07KG5T9W
#FA33FJ% SS*#5%A*
Technical Skills
Programming Languages
ML, AI & Data
Backend & Infrastructure
Frontend
