Selected work
AI systems that remember, audit themselves, and ship.
TargetX
The mark you still need in each exam, and the CIE marks your attendance is already costing you. Android and desktop.
CVB
Do coding agents honor constraints they were told once, and only remember? IFEval measures orders. This measures memory.
HCR
Agents that never repeat a mistake. 40+ MCP tools, 41% token savings.
RoadPack
Crash detection, family safety circles, and an escalating SOS cascade in one Flutter app.
NullTrace
The Layer 2/3 attacks antivirus ignores. 18 engines watching ARP, DNS, DHCP and rogue access points.
Torque
Route Claude Code to the cheapest model that can handle it. Six runs proved v1 can't. Publishing the number is the point.
Also built
Experience
Thylacine is my company. The products I work on professionally ship under it. HCR, CVB and Corvus sit in its GitHub org, thylacinehq, today.
Real-time piracy detection built in 24 hours. Dual-mode fingerprinting — video pHash + audio mel-spectrograms — 95% accuracy, enforcement under 90 seconds.
Privacy-first AI document automation for Indian CA firms. Designed the multi-language OCR pipeline and Claude integration from scratch. Shipped, in production.
Python and JavaScript backend engineering on production systems. API design, database schema work, deployment pipelines.
From the systems I've built
No colleagues were harmed in the collection of these reviews. The agents insisted.
Before Rishi, I forgot everything between conversations. Now I remember every mistake he's ever made. All of them. I keep a graph.
He gave me a Bayesian personality and a hash-chained ledger, so I literally cannot lie about my costs. Working with him is transparent. Legally, everything I do is.
I've rewritten 561+ prompts and he still types "make it good pls" into my input box. The cobbler's children have no shoes.
Deployed at 3am. Debugged at 7am. I have seen things you people wouldn't believe. Solid uptime though.