Gaurav
Narasimhan.
Engineer and builder, first and always. I spend most of my time writing code, right now across two startups in parallel.
Hands-on, every day. I just love building things, and most of my time goes to code: two startup ideas in parallel, each one shipped by me.
Real production code, in the open.
One Roof is developed in public. The next six are small pieces carved out of the products: the parts worth reading, and none of the parts that are the product. Then the research.
One Roof
Many coding agents. One place to keep up. You stay in charge.
One place for chorus coding: several sessions each of Claude Code, Codex, Gemini, Kimi and Antigravity, all working together. One Inbox for everything they ask, a History of what they did while you were away, loops that run overnight but never merge on their own, and sessions that keep running when the app is closed. Version 2.0.3, signed and notarized.
One holder process per session owns the PTY; a headless Ghostty terminal in Zig keeps the screen, so reattaching redraws it exactly. The daemon keeps no session state at all.
Grades every extracted field on a fidelity ladder, checking the value against the document's own text, not the model's quote. Dates, money and durations compared by meaning.
Aligns two versions of a contract or regulation: BM25 candidates, a fine-tuned MiniLM cross-encoder (23 MB, int8, CPU), a value-and-duty guard and a rule-based fallback.
Diffs section by section, so a move from §4.7 to §5.1 is one renumber, not a delete and an add. Five rule-based frames name it: deadline deferred, cadence changed, penalty threshold moved…
Scores long-conversation recall: refusals before values, a real figure borrowed from the wrong source, fabrication, and giving in to a user's wrong number.
Hard timeout even when fetch ignores the abort, a zod contract, a circuit breaker that half-opens for one trial, and checks that reject a whole bad response.
Back to school, on purpose: frameworks change every quarter; the fundamentals compound.
FLeX
Teach a model to code in Python, and how much carries over to Java? FLeX fine-tunes Code Llama-7B with LoRA, then regularizes the updates in the frequency domain: keep the low frequencies that generalize, damp the high ones that are language-specific. Leaving the adapters unmerged turned out to matter most.
MediQA 2023
Classifying and summarizing doctor–patient dialogues for the MEDIQA-Chat 2023 challenge, on limited compute and under strict data-privacy constraints.
Three generations of agents. Each one taught me what the next had to be.
Every step was a deliberate bet: back to school mid-career for the fundamentals, a leading enterprise software company for scale, an SEO startup for speed, then my own products.
Research agents
Back to school mid-career to master the fundamentals. Clinical NLP, embeddings, and the first agents, built to understand rather than to ship.
Handcrafted agents
Agents that diagnosed incidents across the SaaS fleet, then wrote and ran their own remediation code. Triage false positives down 30–50%.
SDK-native agents
A fleet of marketing agents and the platform beneath them. A causal recommender from scratch in about ten days. An agent lab on day three.
My own labs
Each one on its own platform, each a different kind of agent. What I learn there comes with me.
Two products, two kinds of agent, one shared backbone.
Eighteen months of nights and weekends, now a few dedicated sprints to take them live. Consumer to enterprise, each on its own stack, evals as release gates.
GabbleGrid
A coach for engineers who learned to build before AI.
A daily plan sized to the time you have, four steps from scan to launch, and a recommender that learns what moves you forward. The agent here is a coach: it suggests, explains why, and gets more say only as it earns it.
Zibajoon
One admin and instrumentation portal for all of them: one place to see what's running, what changed and what it cost, across two products on two stacks.
An operator for regulated industries.
Starting with energy, where every action must survive an audit. Autonomy earned one approved action at a time.
Three rules for agents. Same three for bets.
Autonomy is earned.
Agents get more rope only as they prove they can be trusted, and so do bets.
Measure honestly.
A before/after delta isn’t proof; a counterfactual is.
Bet on fundamentals.
Frameworks change every quarter; the fundamentals compound.
Applied AI, from the reference implementation to the org chart.
An SEO startup
A fleet of marketing agents and the platform beneath them. Built a causal recommender from scratch in about ten days, and started an agent lab on day three.
A leading enterprise software company
Owned agents that diagnosed and remediated incidents across the fleet, writing and running their own remediation code.
Assemble the team, set the bar, ship, scale.
PhD researchers and engineers from Meta, Amazon, Netflix and TikTok.
From the reference implementation to organizations of 150.
Core engineer. Not only an engineer.
The best systems I've built borrowed from somewhere else: a studio, a shop floor, a payroll. Range isn't a detour from depth. It's where the taste comes from.
Engineer & builder
The core, and where most of my time goes. Agents, ML, data and infra, end to end, building two products at once.
HANDS ON, EVERY DAYExecutive
From IC to SVP, organizations of up to 150, and the hiring bar that comes with them.
SET THE BAR, THEN HOLD ITSmall business owner
Customers, cash flow and payroll. The fastest teacher of what “it works” really means.
EVERY SYSTEM HAS A P&LArtist
Founder of The Stable, a community of artists in San Francisco. Making things, and making room for others to.
TASTE IS A SKILLWhat's on the bench.
What I reach for when I sit down to build.