All Projects.

Autonomous content engine running unattended on sendro.ai. Monitors competitor publications on a schedule, generates B2B blogs and guides through an LLM pipeline, validates every article through a real build gate before any commit, and publishes directly via Git with CDN cache invalidation. Also runs a scheduled social media digest publishing carousels to LinkedIn, X, and Instagram. Lifted sendro.ai organic traffic from single-digit daily clicks to 1,500+.

Migrated sendro.ai from a 350+ page monolithic component system to a Markdown-driven content layer. Applied Ponytail audit principles to eliminate dead abstractions, dead code, and generated scaffolding — reducing the codebase from 400k+ lines to ~65k. Established deterministic heading preservation across content upgrades and set up the publish pipeline now used by Pressroom.

Rust CLI and MCP server that discovers editorial gaps by scanning 30 competitor blogs and 18 knowledge-base sub-sitemaps on every run — no stale caches. Deduplicates against the site's existing content inventory, ranks candidates by topic relevance and source diversity, and outputs suggestions consumed directly by Pressroom.

Background service monitoring Reddit for cold email and sales outreach discussions. Runs on a schedule with configurable subreddit and keyword rules, classifies each post through an AI intent scoring pipeline, tracks leads across a 6-stage qualification pipeline, and dispatches instant Telegram alerts for high-priority signals. Admin panel with CSV export for operational use.

Collects LinkedIn job listings via Apify, then runs a 7-gate qualification pipeline: four free local filters (result count, enterprise blocklist, company size, job description signals) that eliminate poor-fit companies before spending a single API token. Passing companies go through Claude ICP screening, Serper Google search for decision-maker discovery, and a second Claude pass for role verification before inserting into the lead database.

Scrapes Wellfound (AngelList Talent) for SDR/BDR hiring startups using Camoufox headless browser. Built a DataDome bypass technique discovered experimentally: loading the public jobs page first to establish a session cookie before navigating to role-specific pages — no proxy, no IP rotation, no captcha solving needed. Supports two extraction modes (role-page pagination and landing-page fallback). Runs Claude ICP screening and Serper decision-maker discovery per accepted company. Full API cost tracking per run.

Scrapes Naukri.com for SDR/BDR hiring companies using Camoufox running in-process — no external scraper API. Full async pipeline: extract companies via headless browser, dedup by company name, Claude company ICP screening, Serper decision-maker discovery, Claude role verification, PostgreSQL storage. Targets the Indian market where Naukri is the dominant job platform.

Collects job listings from Indeed via Apify actor. Pure collection pipeline — no AI screening, no Serper, no lead gen. Deduplicates by job ID, extracts company domains, and stores into PostgreSQL for downstream processing. Simplest scraper in the stack by design: validate the channel before adding screening cost.

Scrapes startup.jobs by intercepting Algolia XHR responses via Playwright's network event listener instead of DOM parsing — because the page HTML is unfiltered regardless of search query; only the Algolia JSON response carries actual search results. Runs the same Claude ICP screening and Serper decision-maker pipeline as the other scrapers in the stack.

Full async scraper for GetLatka's SaaS company directory. Handles 3,619 paginated pages with a checkpoint/resume system so interrupted runs continue exactly where they left off. Outputs to CSV or PostgreSQL for downstream company registry use.

CLI pipeline that extracts hiring company slugs from Common Crawl CDX index files using binary search on sorted URL keys — finding all companies on a given ATS platform across hundreds of millions of crawled URLs in under 100 seconds. Plugin architecture covers 12 platforms: Greenhouse, Ashby, Lever, BambooHR, Workday, SmartRecruiters, Breezy, iCIMS, Recruitee, Rippling, Personio, and Workable. Includes job-fetching per slug via REST, GraphQL, headless browser, or Sitemap XML depending on the platform. 5-month cross-crawl refresh already completed across all 12.

NLP analytics system enabling non-technical operational teams to query ERP databases in natural language. Built LangGraph stateful workflows covering dynamic SQL generation, automated CSV/XLSX file ingestion, and analytical tool execution chaining. Multi-user session isolation via Redis and PostgreSQL. Real-time delivery through a web dashboard and Telegram bot for operations reporting.

Designed and deployed two private API platforms providing structured digital address resolution services. Built the full backend and API surface, including authentication, rate limiting, and internal automation pipelines.

Designed and built the full website rebranding for Eallisto Labs — eallisto.com. Complete UI rework covering visual identity, layout system, component architecture, and content structure. Delivered a production-grade site reflecting the company's positioning in enterprise AI and NLP tooling.

Experimental local-first agentic framework built during NLP research internship at IIIT Kottayam. Designed a 7-node LangGraph workflow with tool-assisted reasoning, autonomous execution loops, and staged decision branches that skip LLM calls on deterministic steps — reducing inference overhead by ~50%. Integrated 16+ Google Workspace and filesystem APIs via OAuth2. Real-time workflow execution tracked over Server-Sent Events.

Multi-agent system transforming unstructured citizen complaints into structured incidents, danger risk assessments, and emergency response workflows. Qdrant handles vector memory for complaint clustering. Neo4j maps infrastructure topology for graph-based reasoning. Redis Streams dispatch async workers for real-time incident processing. PostGIS provides geospatial context for proximity-based response routing.

Autonomous system monitoring codebases for architectural flaws. Uses semantic search to surface contextually relevant code, graph relationships to trace impact across modules, and an LLM-guided fix generation loop that validates the patch before opening a pull request via the GitHub API.

Built in a 24-hour national hackathon sprint. Reverse-engineered raw automotive CAN bus binary protocol streams, decoded manufacturer-specific message IDs and DBC mappings, and built a real-time analytics dashboard with ML-based accident detection and anti-theft tracking.

Autonomous workflow engine that takes a single natural-language instruction and orchestrates chained tool execution across Google Workspace, calendar, drive, email, and external APIs using stateful LangGraph agent loops.