About
I build agentic AI systems, train foundation models, and recruit the people who ship them. Yeah, all three. It looks scattershot on paper, but it’s the same problem from different angles: finding what breaks, and building what holds.
Most of my work these days is applied machine learning and agentic systems: training foundation models, building long-horizon agent runtimes, and turning messy real-world codebases into things that actually execute, instead of just sounding convincing.
Before all that, I spent years in security research and built a cybersecurity company from the ground up. That background still drives how I think: assume systems fail in strange ways, test the edge cases, and care more about what happens in production than what looks good in a demo. Recruiting came later, but it sharpened the same muscle: recognizing the engineers and technical leaders who can actually build under real pressure.
Professional Journey
June 2025 - Present
New chapter. Moved beyond pure niche headhunting into full-spectrum talent acquisition, supporting hiring across engineering, operations, and business functions for one of the largest water and hygiene technology companies on the planet.
Part of the reason I wanted to work inside an organization at this scale was to understand the complex problems large companies deal with up close: the small operational frictions, talent patterns, communication gaps, and cultural decisions that are easy to miss individually but end up shaping how an organization actually works.
The interesting part: building Global Capability Centers in India from scratch. Org design, talent pipelines, cross-cultural integration, the kind of ambiguous, high-stakes problems where most playbooks break down.
February 2022 - May 2025 · 3 years 4 months
Three years in retained executive search at one of the world's top leadership advisory firms. Placed C-suite and VP-level leaders across aerospace, defense, IT, pharma, and life sciences for Fortune 500 clients. The work went far beyond sourcing: building market maps from scratch, understanding how leadership teams are structured, calibrating with boards and senior stakeholders, assessing executive fit, and approaching people who were not actively looking to move.
Executive search taught me how senior hiring really works when the stakes are high and the candidate pool is small. Most of the job is pattern recognition: knowing which signals matter, which narratives are noise, and how to evaluate whether someone can actually lead inside a specific business context. Built the kind of network you can only build by being useful to very senior people over a long period of time.
2020 - 2022 · 2 years
2016 - 2020 · 4 years
This is where I learned what building actually means. Co-founded a cybersecurity company to replace the legacy signature-based detection that every enterprise was stuck on. We built the real thing: cloud-native architecture, ML-driven anomaly detection, behavioral analysis engines, neural nets for threat classification, fully automated incident response pipelines. Raised $12.8M from Tier-1 VCs and angels who'd been executives at Microsoft, Amazon, and Google.
Scaled from a prototype in a room to a 15-person team closing mid-six-figure enterprise contracts. I set the hiring bar, built the eng culture, ran the GTM motion, learned to translate deep technical roadmaps into language that boards and enterprise buyers actually cared about. Exited through a strategic acquisition that validated the architecture and the bet. Four years of the hardest, most useful education I've ever had.
2015 - Present
Where it started. I taught myself security by taking real systems apart: reverse engineering binaries, writing exploit chains, and finding vulnerabilities in production software used at serious scale. Reported issues across Google, Apple, Meta, Microsoft, Amazon, and Uber under the handle "xedro", and I still help startups, open-source projects, and engineering teams find and fix the failure modes that matter. What stayed with me was not the thrill of breaking things, but the discipline behind it: tracing assumptions, finding edge cases, and understanding why systems fail when they meet reality. That way of thinking still sits underneath how I build models, debug infrastructure, and judge technical work.
My GitHub contribution graph reflects that I still ship code regularly, even as my work has shifted deeper into leadership and talent.
What I Work On
Trumbo is the AI lab I run now, aimed at the failure mode I keep seeing everywhere: fluent answers, brittle execution. We build agentic coding tools, agentic cybersecurity solutions, and Quartz, the model underneath it all. Quartz is the foundation: a proprietary sparse MoE model pretrained on code, security, and synthetic agent trajectories, mid-trained through repo-level context up to 128K, then RL-aligned on coding and multi-step tool use. Lite activates a smaller expert set for reflex inference; Hyper runs plan→reason→verify→synthesize cycles with backtracking when verification fails. An Agent sits on top: Plan mode to map a repo, Act mode to mutate it, with shells, browsers, and tool chains long enough to finish real work. The hosted runtime adds persistent sandboxes, stateful cloud agents, and Sentinel, where 200+ deterministic signals are triaged by models, traced through data flow, and adversarially verified before they ship. The model does the work first, and the explanation comes after it’s been checked. Learn more about Trumbo →
On the research side, the quieter obsession is making agentic models hold under real load, not on the benchmark that flatters them. Most of this doesn’t fit neatly on a résumé, so here’s the honest, slightly too-long list:
- Sparse MoE pretraining: multi-stage curricula on code, security corpora, and synthetic agent trajectories
- Compound reasoning: plan→reason→verify→synthesize loops, with confidence-guided backtracking when a step doesn’t hold up
- Agent runtime: long-horizon tool execution, sandboxed environments, adversarial AppSec verification
- Efficient inference: sparse expert routing, speculative decoding, and production feedback loops that actually feed back
- Calibrated doubt: teaching an agent to know when it doesn’t know, so a fluent answer is an honest answer
- Code as ground truth: using real execution to verify what a coding agent produced, instead of trusting it because it sounds certain
- Data-loop hygiene: tracking how much fresh human signal stays in a training loop versus recycled output, before the echo thins it into noise
- Post-hoc explanation: making reasoning traceable after the fact, so a shipped wrong answer is a bug, not a mystery
- Agent red-teaming: treating a capable agent like a piece of software to break, because the story usually fails at the seam, not the center
- Expert interpretability: peeking at which sparse experts actually fire and why, so a routing decision is a thing you can inspect, not a vibe
Some of this is shipping. Some of it is still being fought. All of it points at the instinct I picked up in security: trust nothing until it’s been checked, and let the explanation come after the work.
Then there’s Glide, the active product on the people side of the house. It came out of years in headhunting, watching capable people get crushed by a hiring pipeline designed to process volume rather than surface excellence. Glide is career intelligence, not another job board: custom models for candidate-job matching and skill verification, prep grounded in real company data, and none of the keyword roulette. Same core instinct as the lab, different side of it: finding what holds under real pressure, and what collapses when conditions shift.
I also advise companies on technical hiring strategy and building leadership teams that can execute on ambitious roadmaps.
How I Think
I start with one question: what’s actually happening here? Not what should be happening, not what the docs say, but what’s actually going on in the system. I ask it the same way whether I’m debugging a training run, evaluating a candidate, or designing a model architecture. Break it into testable pieces. Watch how the parts interact over time. Trust data over narratives. Optimize for resilience, not elegance.
The most interesting problems sit at the edges between domains, where foundation model training meets agent execution, where security research meets autonomous verification, where individual engineering decisions quietly shape whether a system holds under real pressure.
”The task is not so much to see what no one has yet seen, but to think what nobody has yet thought about that which everybody sees.”
Arthur Schopenhauer
The most interesting work lives in overlaps that don’t fit neatly on a résumé. My path from security research to startup to recruiting to ML wasn’t planned. Each phase just built skills that compounded on the last. I follow the questions that hold my attention, especially at the boundary between people and systems, and let the labels catch up later.