Jawahar Prasad
Senior Director of Engineering · Tekion
I build AI dev productivity at scale, and the measurement framework that doesn't rely on AI lineage or surveys.
What I'm building
I'm the architect of ORBIT, Tekion's agentic AI dev-productivity platform, running across thousands of repositories (700+ microservices). About 1,500 people use it across ~20K sessions a month, and the largest group isn't engineering; it's Customer Support.
- DeepContext: graph-augmented retrieval over Jira, Confluence, code, PRs, and support tickets.
- Central: agentic orchestration with sandboxed execution; 145K+ production runs.
- ORBIT Support Agent: answers L1 questions, drafts replies, diagnoses tickets against code and config, suggests fixes, and escalates to engineering. Built on Central and DeepContext; in production and rolling out to all customers.
- CPI (Composite Productivity Index): productivity measurement that doesn't depend on AI lineage, surveys, or tool-usage counts. Paper in development.
- PACE: a Jira effort-estimation model I built. 85% of estimates land within one size bucket. It ignores developer story points, which were teaching it to repeat chronic under-estimation.
- Grit: a support-ticket triage model I built. Nine TF-IDF + logistic-regression classifiers, one per routing decision, each with a calibrated confidence gate. It beat a fine-tuned 566M-parameter embedding model on held-out tickets, so the lightweight model shipped.
- AIR: AI incident response, in early preview.
ORBIT replaced PDAF (Persona-Driven AI Framework), which I shipped in 2025. Augment Code featured it in December 2025: merge time 3 days to 1 day, MR volume +21% at the same headcount.
How we measure it
We track whether AI code review actually changes what ships. In one quarter, the share of high-priority review findings fixed before merge rose from 57% to 77%, at about $5 per PR.
Selected work
- LeadDev: How Tekion enabled 1,300 employees to work with agents
- Augment Code customer feature, Dec 2025: 60% faster merges under PDAF
- CPI: Measuring AI Impact on Engineering Productivity Without Lineage or Surveys (paper in development)
Open source
- TokenDog : a local-only watchdog for Claude Code token spend. It prices every turn and flags sessions worth closing, and it's upfront about what it doesn't reduce.
Experience
Twenty-one years building large-scale platforms, about 16 of them in ML, search, shopping, and AI. Yahoo!, Jabong, Flipkart (doubled ads revenue; Best People Manager 2018), Amazon (Recommendations and Personalization). Founding team of what became Whatfix (2010); left after 13 months, before product-market fit.
Bay Area · originally from Bangalore.
Elsewhere
Get future posts
No newsletter yet. I'm publishing on LinkedIn for now. Want a heads-up when I launch one? Click to reveal my email address.