The professional side_

M.S. — Statistics & Data Science

New Brunswick, NJ, USA

# relevant coursework

B.Tech — Computer Engineering

Jaipur, India

# relevant coursework

An AI agent that gets small B2B businesses paid: it decides when, how, and on which channel to chase each overdue invoice — email, SMS, or a live two-way AI phone call — drafts the message in the business’s own voice with Gemini, and only sends after a deterministic compliance gate clears it.

• ▸A seven-stage agent pipeline with one safety authority: source adapters normalize invoices into a canonical schema (derived fields like days-overdue are always recomputed, never trusted), then ingest → strategy → LLM judgment → drafting → compliance gate → gated send → reconcile — every decision written to an append-only execution log that doubles as an explainable activity feed.
• ▸The LLM is deliberately not trusted: Gemini refines tone and timing inside a hard clamp (it can never flip a “hold” into a “send”), and 15 deterministic rules gate every message — legal-threat / consumer-debt / dispute detection, contact-frequency caps, and a non-custodial payment design where the real Stripe link is resolved only after the gate (any hallucinated URL escalates to a human).
• ▸The voice channel is a compliance system, not just TTS: live calls via Retell (ElevenLabs voices) open with a gate-enforced verbatim AI disclosure, with per-debtor consent records, an automatic do-not-call registry, local-clock call-hour windows, a never-double-dial ledger, and HMAC-verified webhooks that turn ended calls into audited transcript artifacts.
• ▸271 hermetic tests (zero network), including red-team suites that fire adversarial drafts, call scripts, and transcript injections at the gate — proving the core invariant: a mis-wired pipeline still cannot send a message or ring a phone.

Turns cold job applications into vouched, warm intros. It verifies a candidate’s real competence from their GitHub, portfolio, and a short AI work-sample, generates a “safe-to-refer” brief, and routes it to the right insider at a target company — so an employee can vouch for a stranger without staking their reputation.

• ▸Next.js 16 / React 19 / TypeScript on Supabase (Postgres + row-level security). A verification engine pulls live evidence from the GitHub API and the candidate’s portfolio, and Gemini turns it into an evidence-backed competence score plus a per-role safe-to-refer brief.
• ▸The differentiator — a Tier-3 “AI work-sample” that auto-generates 2–3 questions anchored to the candidate’s own repositories, grades the typed answers 0–100, and folds that live signal into the score, so it can’t be faked by a stranger or a stock LLM prompt.
• ▸A fully automated two-sided loop: match the right insider via Hunter (role-scored) → email the underwritten brief via Gmail → accept or decline from a single-use, rate-limited public link with no login, feeding an outcome-ranked referrer-reputation model. Employers post a role, get an AI-verified shortlist, and pay a Stripe success fee on hire.
• ▸Production-hardened: an SSRF-hardened portfolio fetcher (rejects private / loopback / cloud-metadata ranges and re-validates every redirect), résumés generated by constraining the LLM to structured JSON inside a fixed LaTeX→PDF template, zod input validation, per-endpoint rate limiting, and graceful mock layers so the whole app runs end-to-end with zero API keys.

A Retrieval-Augmented Generation system engineered for accuracy and speed — LLaMA 3 + LangChain + FAISS, served in production, not a notebook demo.

• ▸+70% response accuracy over a naive baseline by tightening retrieval and prompt construction.
• ▸Low-latency vector search with FAISS — query time cut 90ms → 30ms (~3× faster).
• ▸Served behind a FastAPI interface with CI/CD via GitHub Actions and automated performance tracking.

A recommender that predicts your next purchase by fine-tuning open LLMs on real e-commerce behavior instead of classic collaborative filtering.

• ▸Fine-tuned Mistral-7B and TinyLlama on the Amazon Reviews dataset using LoRA / QLoRA for efficient training on modest hardware.
• ▸98%+ next-purchase prediction accuracy via novel prompting over product and review text.
• ▸Turns unstructured product language into precise, personalized recommendations.

A rigorous head-to-head of RAG architectures across accuracy, coherence, and latency — with a verdict, not just vibes.

• ▸Benchmarked naive, hybrid, and re-ranked RAG end-to-end; hybrid retrieval won the accuracy/latency trade-off.
• ▸Built standardized evaluation metrics so retrieval strategies could be compared fairly.
• ▸Quantified how retrieval choices cut hallucinations while preserving contextual relevance.

A real-time analytics dashboard that turns daily public-safety data into an interactive map of campus crime hotspots.

• ▸Automated Selenium pipeline scrapes fresh incident data from SpotCrime every day.
• ▸Interactive Plotly charts and geospatial maps surface where and when incidents cluster.
• ▸Blends data engineering, visualization, and automation into a live public-safety tool.