Mohit Asthana
I carried a quota and built internal value tooling for AE teams at Salesforce. Now I build and run production AI agent systems solo, full time.
San Francisco Bay Area
What I've built
A production content pipeline, not a demo
A Next.js console over Supabase Postgres, with a pg-boss job queue running 50+ handlers: idea generation, captioning, carousel and short-form video rendering, voiceover, newsletter drafting, publishing, and metrics polling. This runs my own media business end to end, not a portfolio toy.
View the ai-marketing-os repo →A Telegram control plane, laptop off
Two bots on a small cloud droplet carry every approval, daily content triage, and feedback to my phone. Nothing publishes without a human tap, and the approve-to-publish loop runs entirely in the cloud. Batching, silent tiers, and quiet hours cut notification load from around 60 pings a morning to 3 to 5 that actually matter.
View the human-in-the-loop doc →Cost discipline as a design constraint
Cheap models by default through OpenRouter, escalating only where brand voice precision earns it. Paid renders sit behind an explicit approval tap. Auto-generated shorts are capped at 2 a day. Total infrastructure runs about $12 a month across two small droplets, plus metered AI usage.
View the cost-controls doc →Postmortems when it breaks
My pipeline once stalled silently for about 12 days on an unapplied database migration plus one unset environment variable on a new host. I root-caused it, fixed it, instrumented it, and wrote it up, the same way I would for a team, even though there was no one else to report to.
View the postmortems repo →An AI calculator my fellow sellers adopted
At Salesforce I built an interactive business case calculator with Claude: 7 value drivers, a 3-year financial model, and a printable one-pager. Other sellers picked it up on their own and used it on live deals; one teammate called it "insanely helpful for winning a big bet deal." Building a tool is easy. Getting salespeople to adopt one voluntarily is the hard part.
See the tool → See the shoutout →User-facing products, shipped live
Agent pipelines aren't the only artifact I ship. Using Lovable plus my own domain, export, and pricing work, I've put a free affordability calculator live at caniaffordthisproperty.com and built the two tools that run my consulting sales motion: a prompt playbook and a proposal generator. Screenshots below.
See the products →Agents with real permissions, behind a real gate
Hermes is my production agent control plane, running on a $12-a-month droplet: 214 tools across 7 MCP servers. External writes (spend, sends, live changes) are designed to route through a root-owned approval gate with out-of-band Telegram approval codes. An independent sentinel watchdog checks the system every 15 minutes, and code changes ship through a PR-only lane with no GitHub token on the box. Agents with real permissions need real guardrails.
An agentic SEO system that caught its own defect
For my agency, SalesOS Labs, I shipped a 66-article topic cluster in about a week: full schema markup (Article, FAQPage, Organization, Person), llms.txt for answer engines, and a 105+ URL sitemap with real lastmod on every URL. Then I diagnosed a structural defect, 100% of internal links pointing at non-canonical URLs, and fixed it at the build level: 0 to 1,213 canonical internal links, with a validator so it can't regress.
See the live blog →VerifiedOTE, a live SaaS we run
Co-built with a business partner and launched as a live product. Stripe live-mode checkout is verified end to end in production, and the daily digest email has run 13 consecutive days at 100% delivery with zero bounces. We watch the boring numbers because the boring numbers are the product.
See it live →Tools other people run their work on
I built and operate a newsletter production pipeline (autowriter, issue configurator, publish-ready paste for Beehiiv) that a human editor uses to draft and ship every issue of a curation digest. Content selection is driven by a frozen 0-22 scoring model over archived performance data. Personal automation is easy; a tool another person relies on to ship has to actually hold up.
See the brand it publishes for →The system, live
Shipped products, built with Lovable
Not everything needs a job queue. When the right artifact is a user-facing product, I use Lovable to go from prompt to deployed app in days, then wire in the custom domain, exports, and pricing logic myself. One of these is live on its own domain for a real audience; the other two run my consulting practice's sales motion.
Constraints and self-healing
An autonomous system only earns trust when it is constrained and observable, not just clever. MohitOS hard-caps spend and gates every paid or public action behind an explicit human tap, then reports its own health on one strip instead of scattering signals across dashboards nobody checks. When something is actually broken, the system says so out loud: a startup wizard checks its own 31-item configuration and prints the exact command to fix whatever is missing, and a watchdog kills and restarts any process that goes quiet instead of limping along undetected.
How I work
Every real incident gets a writeup: what broke, why, and what changed. Four are public in the postmortems repo, including a silent 12-day pipeline stall and a database quota I found the hard way.
Cheap models handle the high-volume work by default. Stronger models get called in only where the output quality actually depends on it.
Agents generate everything. Nothing that costs money or reaches an audience goes out without a human tap first.
I put the system in front of real usage early, let it break in small ways, and fix what the breakage actually teaches me instead of over-designing upfront.
What colleagues say
"Mohit is a motivated self starter with a reputation for quality work and a track record of success. He is resourceful, dependable and thinks outside the box about how to solve complex problems. [...] Mohit is also organized and concise which are rare qualities that I value in my teammates."
Chaz Van de Motter, LinkedIn recommendation, public on my profile
And two straight from Slack, as they were sent.
Why forward deployed / AI ops
Carrying a quota taught me how to run discovery, manage a room of stakeholders with different incentives, and hunt down new business instead of waiting for it to land in my lap. I hit my new business quota at Salesforce the same way I ship a system: work the real problem until it closes. Forward deployed work is the same job pointed at a different artifact: sit with the customer, understand what they actually need, then go build it. At Salesforce I did the quota-carrying discovery and stakeholder half. Now I run the build half full time. Both halves are proven; this is the role that points them at the same problem.