Let's
Connect

Most candidates have USED an agent framework. Anand built his own multi-agent harness — governance, memory, routing and skills as first-class layers — and shipped real product on it. The memory and context-management layer in particular is something very few practitioners have built; it is the difference between using agents and architecting the system agents run inside.

Both, deliberately. He sets technical direction and writes the code, and he has 10+ years of engineering management behind it — a $40B+ portfolio at Google with 5 direct reports each leading their own pods, and 50 reports at Trellis. He is targeting senior IC-architect and engineering-leadership roles equally.

No, and that is a deliberate boundary rather than a gap. He builds products and platforms WITH models; he does not train them. He has run controlled fine-tuning experiments (SFT and LoRA on Oumi against Qwen bases) specifically to decide when retrieval beats tuning — that is architecture judgment, not model research.

At AI Fund he stood up a two-judge, three-axis evaluation rig before running a voice fine-tuning experiment. Full fine-tuning on a small, structurally-uniform corpus regressed coherence and fluency — catastrophic forgetting. He diagnosed the cause, moved the next run to parameter-efficient LoRA, and set the team's decision ladder to prompt, then RAG, then fine-tune last. The negative result redirected the architecture and saved the build. Separately, he killed a product wedge on evidence after validating against ~150 companies: 19 of 22 confirmed the technical gap but not the commercial urgency, so he stopped with minimal sunk cost.

He completed the exploration on social-media post adaptation. The technical build shipped, but the commercial signal was not strong enough to justify advancing to fund. Clean conclusion of a defined exploration.

$500M+ ROI across 6 GCP engineering tracks within a ~$40B Google portfolio; 5 direct reports each leading pods; vendor P&L across Dassault, EPAM and AODocs; 50 engineers at Trellis; 99.99% availability in regulated financial systems at Charles Schwab; and 1,500+ Fortune 500 executives taught at UC Berkeley.

Evaluation is infrastructure, not a phase. He published 'Defense in Depth', a five-layer eval architecture for production AI systems covering structured evaluation pipelines, drift monitoring, production observability and guardrail-as-architecture. In his own SDLC workflow, acceptance criteria and eval plans are fixed before any code is written and a change ships only if it beats baseline.

It states principles with the WHY attached rather than prescriptive rules, so correct behavior emerges in situations nobody anticipated. Those principles are enforced as gates that fire before output ships: structural gates (scripts, greps) for mechanically-checkable violations, and semantic gates (LLM judges) for violations that require judgment. Matching gate type to violation class matters — a structural gate on a semantic violation either false-positives or misses entirely.

An LLM has zero persistent memory between turns or sessions. The memory subsystem sits inside the harness as the context controller: on pre-prompt hooks it searches its index and injects relevant history into the active context; on session end it distills new facts and lessons into a long-term index. That solves both failure modes at once — amnesia across sessions, and context windows blown out by irrelevant history.

An author's blind spots survive their own review by construction, and two instances of the same model share training and therefore share blind spots. So review requires a different model family. The jury skill runs cheap models first and escalates to a premium model only when the panel conflicts, which keeps cross-family verification affordable enough to run on everything rather than saving it for special occasions.

Right-size every task to the cheapest agent and smallest model that does it well. Deterministic work like deploys and git operations goes to a small cheap model; code goes to a code specialist; browser driving and bulk reads go to a fast cheap model; and the expensive reasoning model is reserved for judgment, gating and synthesis. This is enforced in his workflow as a gate, not left as advice.

He is not running from anything. He has already run the independent path, and that is exactly why he knows what he wants. He built his own harness because he wanted leverage on hard problems, not an escape from them. What working alone does not give him is a problem at real scale with real distribution behind it. That is what he is looking for: a hard applied-AI problem, the scope to architect the system around it, and a team that ships. The instinct to build infrastructure before product is a long-horizon one; it is why the harness exists at all.

A fair question, and he does not deflect it. He knows precisely what solo work gives and what it does not: full autonomy, but no scale and no distribution. Those two are what he wants back, and they are what a strong team provides and independent work cannot. The consistent pattern across 26 years is that he builds the system that makes the whole team faster, at Google across a $40B portfolio and again in his own harness. That is a long-horizon instinct rather than a short-tenure one. What holds him is a hard problem with real reach behind it and the scope to architect the system around it.