Initializing Arc Reactor
0%
From AI-powered applications to full-stack products. Each project taught me something new about building software that matters.
A five-sector daily news dashboard that runs on zero infrastructure: a scheduled GitHub Action pulls ~71 RSS feeds, dedupes and ranks the survivors, and Netlify redeploys a static site on every commit.
A professional-grade Pomodoro timer built in Flutter with an AI coach: natural-language session input, Gemini-powered productivity insights, and ambient soundscapes for staying in flow.
Compliance analysts auditing dense policy documents get grounded, evaluable answers instead of a black-box chatbot — a self-correcting CRAG + Self-RAG pipeline measurably improved retrieval precision and answer faithfulness against a Ragas golden dataset.
PMs who spend hours each week triaging feedback scattered across Slack, support exports, and documents get a weekly, citation-verified brief instead — a 5-agent LangGraph pipeline where every quote is copied verbatim from source and independently re-verified before it ships.
Knowledge workers who need trustworthy answers from long internal PDFs get every response with the source page number and a relevance score attached — a multi-agent RAG system built on LangGraph orchestration and an event-driven cloud ingestion pipeline.
A dynamic utility designed for Product Managers to manage probabilistic AI systems. Leverages background agents to automate the synthesis, prototyping, and testing of new product features.
An AI-powered job search pipeline that automates lead generation and evaluation. Solves the fragmentation of job boards by aggregating sources, scoring leads via LLMs, and refining accuracy through feedback loops.
A multi-agent system that simulates an Agile squad to autonomously refine feature requests into comprehensive, estimated tickets — streamlining the product planning phase.
Developers stuck turning "it looks broken in this recording" into a code fix get an assistant that watches the screen recording and reads the code at the same time — built end-to-end during Google DeepMind's one-week Vibe Code with Gemini 3 Pro sprint.
SMB recruiters who screen hundreds of resumes by hand — and lose good candidates to formatting, not fit — get autonomous parsing, scoring, and ranking with a reasoning trace behind every score. Shipped as the Phase-1 "Screener" agent in a roadmap that scales the same core into scheduling, sourcing, and bias-audit agents.
Engineering teams that need to stress-test AI agents before real production data exists get an autonomous multi-agent pipeline that fabricates structurally valid, adversarially diverse synthetic datasets — cutting dev-cycle time 10–20x and infrastructure cost up to 80% versus hand-built test data.
An adaptive, AI-powered study mentor for GATE aspirants that personalizes the learning journey through real-time adjustments and proactive nudges.
A production-grade, modular RAG application designed to handle complex information retrieval across diverse data formats using intelligent query routing.
A full-stack travel booking platform that streamlines trip planning with dynamic content management, secure payments, and AI-driven advisory.
An end-to-end machine learning application that predicts critical machine failures in real-time, enabling proactive intervention in industrial settings.
An AI-powered health assistant that democratizes access to complex medical information by analyzing reports and summarizing dense web articles.
A comprehensive career optimization tool that maximizes candidate success by using AI to align resume bullet points directly with target job descriptions.
An intelligent learning companion that accelerates professional development by creating personalized, actionable roadmaps based on user goals.
A secure, full-stack journaling application that helps users cultivate self-awareness by tracking mental health trends and analyzing emotional states.
An AI-powered productivity tool that transforms unstructured daily logs and handwritten notes into formal, structured Weekly Progress Reports for university compliance.
SMB recruiters who screen hundreds of resumes by hand — and lose good candidates to formatting, not fit — get autonomous parsing, scoring, and ranking with a reasoning trace behind every score. Shipped as the Phase-1 "Screener" agent in a roadmap that scales the same core into scheduling, sourcing, and bias-audit agents.
Recruiters at SMBs spend 20+ hours a week screening resumes by hand. The persona behind the product — "Sarah", a solo talent-acquisition manager at a 50–150 person company — gets 500+ applications within 24 hours of posting a remote role, driven by one-click apply and AI resume spam. She burns 3–4 hours a day reviewing PDFs; by the 50th resume, decision fatigue sets in and hidden gems get rejected. Screening 200 CVs takes ~5 days, while candidates wait 14 days for a response — or get ghosted.
An event-driven serverless agent (n8n + Gemini 2.5 Flash) replicates recruiter decision-making: webhook ingestion, OCR perception of the resume, then a gap analysis against the job description under a "Senior Talent Acquisition" persona, producing a weighted 0–100 score (40% skills, 30% experience, 30% soft skills) as schema-enforced JSON with summary, pros, and cons. The scoring prompt took three documented iterations — a chain-of-thought rubric finally caught nuances like "used React for 1 year" vs "Senior React Developer". When CORS and 15–20s AI latency broke the synchronous flow, the architecture pivoted to fire-and-forget: upload to Supabase Storage, process in the background, and push scores to the Kanban UI over Supabase Realtime, so the app feels instant. The split-screen candidate view came from user research: recruiters will not trust a black-box score unless the evidence sits beside it.
Shipped as the Phase-1 "Screener" agent with a live demo and a full product paper trail — PRD, metrics & roadmap doc, and a process/decision log. Defined KPI targets: time-to-shortlist cut from 5 days to under 1 hour, interview conversion lifted from 15% to 40%+, and cost-per-hire pushed from the $4,700 industry average toward $500. The roadmap scales the same agent core into Coordinator (auto-scheduling), Hunter (autonomous sourcing), and Compliance (bias-audit) phases.
As the product owner, defined and tracked three success metrics end-to-end: funnel analytics on the Job Post Creation flow to catch drop-off, a retention dashboard for daily recruiter activity and platform stickiness, and a North Star metric of reduction in average screening time per candidate.
Started the resume-parsing pipeline on GPT-4o, then swapped to Gemini 2.0/2.5 Flash once the n8n workflow was underway — OCR-heavy resume parsing needs a large context window and low latency, and at the volume an SMB screening 500+ resumes actually sees, GPT-4o's per-call cost stops penciling out against Gemini Flash's price-to-performance.
Cut a custom Python backend in favor of building the agent workflow visually in n8n (Webhook → PDF parse → Gemini → JSON parse → Supabase) — as a solo founder, hand-building routing, auth, and state management would have consumed most of the build time; the visual graph made it possible to debug the agent's reasoning steps without restarting a server.