I take the process quietly bleeding a business time or money, design the agentic system around it, and ship it live. Voice agents that answer the phone and book the appointment. Software that reads a contract and writes the offer. A brain that runs an entire sales floor from Slack. Twenty-plus systems shipped across five countries, many still running in production.
Evaluating and calibrating the newest coding models against detailed rubrics, and co-authoring AI research. I can't name them here, but it keeps what I ship current with where the models actually are.
I came up as an engineer at PhonePe, India's largest payments platform, building for hundreds of millions of users. Reliability at that scale is the bar I hold my own work to.
I run a small AI automation studio and ship as the builder, hands on the keyboard. I own the outcome, sit in production, and fix it until it holds under real load.
Across construction, healthcare, finance, sports, and services. Client names are held back under confidentiality. The systems and the numbers are real.
Every client, pipeline, invoice, SOP, and outreach sequence in my studio runs through one Claude Code-native system I built. It is why a two-person team ships like ten. I am now packaging it into a plugin marketplace and a SaaS so other agencies can run on the same rails.
Drop a business card, a voice memo, or a conference recap into Slack, or forward an email, and it extracts the contacts, enriches the leads, writes the CRM record and a per-contact dossier, and drafts the follow-up in the founder's own voice. Three pipelines on n8n, roughly forty-seven nodes.
It reads a company's own ERP, accounting file, bank statements, and spreadsheets read-only, returns one trustworthy set of numbers with every figure traceable to its source, and catches the wrong one before it gets filed.
Built with a Dubai physician. Voice booking on VAPI, Retell, and LiveKit, an AI chat agent, patient scheduling, medical-query handling, consultation recording, and document intelligence, all handling real patient interactions.
Spanish-first over WhatsApp. It books, reschedules, and cancels straight into the calendar, handles the local insurers, transcribes patient voice notes, redacts protected health information, and routes emergencies after hours, on a real clinic number.
HIPAA-compliant. It bridges the pharmacy's software to a voice agent that pulls the patient, prescription, and insurance data, places the call, verifies identity, delivers the message, and captures the outcome across seven paths and twenty-plus edge cases.
Website visitors book racquet-sport courts over both text and live voice from a single script tag. A backend proxy protects the key and stays the single source of truth for config, so it drops onto any site cleanly.
An inbound and outbound voice agent for a wellness clinic, wired to a GoHighLevel CRM with quiz-based segmentation, intake forms, payment sync, and compliance handling. It replaced manual phone reception end to end.
I automated an inbox-setup company's entire fulfillment flow and built them a SaaS dashboard for full operational visibility. Fulfillment time dropped fivefold and they added five figures in monthly recurring revenue within weeks, delivered in under four weeks.
A prospect requests a demo call, standard or custom, and can upload their own FAQ, product, and flow documents to spin up a personalized voice demo on the spot. I built the product and the demo funnel behind it.
I mapped the current-state workflows across every service line of an engineering firm, then automated their fee-proposal process across seven templates and designed HubSpot dashboards giving leadership real-time visibility into pipeline, revenue, and utilization.
For a company running about 2,600 live deals and recurring-payment automations. I built a revenue-recognition engine and the deal-type and lost-reason model as additive, non-destructive changes, recognizing revenue in the month the service is delivered.
Four scrapers pull founder and startup events across twenty-five US metros. The pipeline dedupes, scores each one zero to a hundred with an LLM, enriches to clean structured data, and mails out what clears the bar. One run took 1,578 in and shipped 410. Backed by 57 tests.
Point it at any B2B list. It grades every account against an ideal-customer profile, enriches the survivors with public signals from their site, hiring pages, maps, and news, then scores and tiers them A through F. One run took 7,842 accounts to a clean set.
It scrapes buying signals before anything else, a new exec in role, headcount growth, recent funding, live hiring pages, then buckets accounts by pain, writes copy per bucket, and extracts emails last. Free-first by design and hard-capped at five dollars of scraping per three-to-four-thousand-lead run.
An always-on pipeline that watches eleven feeds, writes opinionated takes with matching imagery, and posts them for a human to approve before anything ships. A sixty-five-node workflow, plus deliverability-tuned outbound that gets read.
I studied his seven process walkthroughs and turned a twenty-five to thirty-five minute per-property offer process into a five-minute conversation. Commands run his offer calculator against his own workbook, log the deal, fill the Colorado purchase-agreement PDF field by field, and draft the agent email.
Land feasibility to auction war-room for residential developers. It turns raw site data and regulatory constraints into a go or no-go call inside the Indian RERA workflow, a place no existing tool reaches. Validated live on real auction lots, built with my engineering team.
A hundred-thousand-line trading app. A fifty-six-finding audit, every finding tied to a file and a line across six rounds of adversarial verification, then around twenty-nine fixes with the suite green at 356 of 356. One catch: a fee bug that would have stopped any deal over seventy million dollars from reaching a human.
An internal store where a 250-person team browse, submit, and review their own AI tools under role-gated approval. I own the safety layer, a review harness and hard rules on every pull request, so non-engineers ship internal tools without breaking production.
I grade and calibrate the newest coding models against detailed personality and instruction-following rubrics, across dozens of task sets, and co-author AI research alongside it. I stay close to how these systems actually behave, then build on top of them. The labs go unnamed by agreement.
I started out as an engineer at PhonePe, India's largest payments platform, where a bug is not a bad demo, it is someone's money. That is where I learned to build things that hold.
I trained at Delhi Technological University and now run a small AI automation studio. I do the building. I take a real business problem, design the agentic system around it, ship it into production, watch it break, and fix it until it holds under load. No prototypes handed off to disappear.
On the side I work with frontier AI labs on model evaluation and research, which keeps me at the edge of what these systems can actually do. What I ship reflects where the models are today.
The brief is never the tech. It is the hour someone wastes or the deal that leaks. I automate the whole loop, then measure against that.
I do not hand back a prototype and vanish. I run it in production, watch it break under real load, and fix it until it holds.
I evaluate and research these models for a living, so I build for what this generation can actually do today.
A tight engineering team and a stack that lets one person do the work of five. That is how twenty systems stay live at once.
The tools I reach for. The list moves as the models do.
Also working across VAPI, Retell, LiveKit, Deepgram, GoHighLevel, Attio, Instantly, and Apify.
The fastest path is a short call. Bring the problem. I will tell you honestly whether it is worth building and exactly how I would ship it.
