Rajamohan Jabbala
CTO who builds AI agents that act in the real world — calling, messaging and deciding under rules no prompt can override — and the organisations that ship them.
CTO, Continuous Learning Systems (CLS++) · Ex-Amazon · IIT Madras
“An agent that calls a borrower has to act under rules no prompt can override. That is why the governor is code, not a model, and why every refusal lands on the ledger next to every send.”
What I own at CLS++
CTO for Continuous Learning Systems (CLS++), the agentic collections platform for lenders. I own architecture, engineering and delivery, and write production Python. The platform is multi-tenant and region-pinned, runs SMS, WhatsApp and voice through Exotel and Twilio behind one carrier interface, and runs on the lender's model or ours. Every decision and every refusal lands on a SHA-256 hash-chained ledger the lender can re-verify offline. Right now I am recruiting the first five design-partner lenders; every pilot is measured against a holdout the lender chooses.
Before CLS++
Chief Technology Officer
Bonito · Dec 2022 – Mar 2025- Series D, ~600 people. Reported to the CEO with board representation; owned the AI platform, the engineering organisation and a ₹20 Cr annual technology budget.
- Built the AI engineering organisation from 5 to 50+ in 18 months and shipped six production AI systems to 50,000+ users.
- One Design Platform: design cycle 95 → 38 days, designer productivity 1 → 2.5x.
- Suchi, a GPT-4 customer-support agent: ₹7 → ₹2 per call, 4,000+ calls a day, 85%+ CSAT.
- Project NCS lifted GTN 7 points, 35 → 42. Project Raven took support tickets from 22 a week to near zero.
- LeadSquared across every call centre: 1.59M leads, zero leakage, 2x sales productivity (published case study).
VP Engineering
ClearTax.in · Oct 2021 – Dec 2022- Series C fintech processing ~10% of India's business invoices.
- GST reconciliation for 500,000+ businesses: 25 minutes → under 10 seconds, with ML anomaly detection on Spark.
- 40% fewer production incidents; 60% lower support cost.
Senior Software Development Manager
Amazon · Oct 2017 – Oct 2021- Led Last-Mile Execution Systems for a delivery network serving 100M+ customers.
- ML capacity planning across 65,000 delivery agents behind a $100M+ a year savings programme.
- Alexa Video's first US transactional video-on-demand: $12M in year one, 3.5% conversion within two weeks.
- Fulfilment at 459K packages a day on average, 1.5M a day at peak.
Co-Founder and Director of Engineering
Aurospaces · 2014 – 2017- A services marketplace from zero to 100,000 customers and 30,000 providers across 22 categories and four cities.
- Built and led the engineering organisation to 40 engineers.
- Invented and patented Elastic Neighbourhood for supply-demand optimisation.
Earlier: Velankani Software, Credit Suisse First Boston, Wipro, Sun Microsystems, Ericsson.
M.Tech, Computer Engineering
IIT Madras
M.S, Applied Statistics
Osmania University
Patents and publications
Elastic Neighbourhood for positioning of display units
Patent · granted
ZPML — privacy-preserving ML retrieval
Patent · provisional, filed
Raster Snap Function for Vectorization of Bitmap Images
Publication · IEEE
Cutlist Optimization Ensemble Algorithm
Publication
Selected work
CLS++
Agentic collections · now
Seven agents read every borrower reply, rank the book by provisioning at stake, call and message inside the 08:00–19:00 RBI window, and learn from every contact. The governor that allows or refuses every send is code, not a model.
LeadSquared × Bonito case study
Published proof
The CRM and 13-stage customer journey I put in across Bonito's call centres as CTO, in the vendor's words: 1.59M leads managed, zero lead leakage, a 500-agent auto-dialer.
Agent Transport Protocol
Open source · Rust
A five-layer protocol for multi-agent communication, built at CLS++ in 2025 and released as open source.
Agent memory
Earlier at CLS++ · 2025
Lineage-tracked memory for agents: versioned beliefs, WAL-backed persistence, crash recovery, shipped as an MCP server and a Chrome extension. Privacy-preserving retrieval with homomorphic encryption and additive secret sharing is the 2026 provisional patent.
Ten minutes from now, your book is ranked.
Sign up, connect the book, paste two URLs, send yourself a test. One product line is enough to start, measured against a holdout you choose.