Aman Shetty
I run the technical and product engine at Chronomation. Vision, architecture, the code that has to hold up when it matters. Made 21 enterprise-level projects to date.
Five things, end-to-end.
- Agent architecture. Multi-tenant WhatsApp lead-qualification bots with per-client prompts, scenarios, brand voice, and a single Python process serving many tenants. The Chronomation backend (codename chronomation) is mine.
- Memory infrastructure. A 13-stage hybrid retrieval pipeline with 6-term live decay, benchmarked at 93/95 top-3 on 1,547 files. Graph memory for character-driven RPGs. Both built and working.
- Tax automation. Chronos Notice — an Electron+Python desktop app that reads GST and IT notices, extracts structured fields, and never sends client data to a cloud we don't control.
- Founder-voice AI. Custom personas per business, tuned through real conversations against the real model. Not generic chatbots — agents that sound like the founder wrote them.
- This site. The redesign you're reading, the schema graph, the JSON-LD across every page, the cache-bust convention.
Ship the product, then sell the product.
Most studios pitch first and build later. We do it the other way around. By the time we tell a prospect "this works," there's already a running instance we can show them. That bias toward shipping is also why I don't gold-plate — phase 1 is scaffolding, phase 2 fills two functions, every phase has a "done when" criterion. If you can't verify it, you didn't finish it.
I'm brutally honest about scope. If a feature isn't in the MVP, I'll say so on the call instead of agreeing and quietly cutting it later. If a prompt change makes the bot worse, I'll roll it back the same hour. Founder-led means the person debugging at 03 AM is the same person who quoted the timeline.
And the rule I keep coming back to: never call the model unless you have to. Cheapest path first — regex, then heuristics, then the LLM. A spam burst should spend zero tokens. A confirmed opt-out should spend zero tokens. The architecture rewards restraint.
Direct lines. No gatekeepers.
I read every message that comes in through the channels below. If you want the technical conversation, LinkedIn or email is the right door. If it's urgent, Aryan picks up WhatsApp faster than I do. If you want the texture before the conversation — what I type on lives at /uses and what's eating my hours this month is on /now.
Three live tracks.
The Chronomation site you're on. Iterative refinement — JSON-LD breadth, founder profile pages, cubic-bezier consolidation, schema breadth. Each round has a critic agent and a regression agent. The version string at the bottom of every CSS link tells you which round shipped.
The cross-device file finder. Natural-language search across 1,547 files with a 13-stage retrieval pipeline. Currently tuning the decay terms — the 6-term live decay model is where most of the recall gains live.
Memory benchmarking. Writing up the methodology behind the 93/95 top-3 number, including the anti-patterns. The essay lands in /writing when round 4 ships.