Elamaran Elangovan

The through-line

The model proposes. It never decides.

That sentence is in the source of three of these projects, and it is the whole argument. A language model is a fast, fallible generator — useful, and not something to hand a database password to. So the interesting engineering is never the prompt. It is the machinery around the model that makes its output checkable: a registry it cannot invent past, a validator written in ordinary code, a score it isn't allowed to compute for itself, a sandbox where the first mistake lands.

Guidance that would otherwise rot silently repairs itself — and proves the repair before anyone depends on it.

Remove the model from Kerb entirely and it still authors correct, conservative plans. In Agamemnon the model gets a vote, and that vote can only make a decision stricter, never looser. That isn't a fallback bolted on at the end. It's the design.

Work

Agamemnon · Sept 2026TypeScript · Convex · Nebius

A safety layer between an AI agent and your production database.

So an autonomous agent can't quietly delete your company — even when it's fully authorised. In 2026 researchers recorded at least nine cases of an agent wiping live company data on its own. Every time, the agent was allowed to do it, so no alarm ever fired.

Agamemnon takes the database password away from the agent. Every attempted write becomes a proposal: the request and its audit record are saved in one transaction, then Agamemnon counts — with its own database access — exactly how many rows would actually be affected. Not the number the agent claims. Plain readable rules decide allow / needs-approval / block; a tuned model gives a second opinion that can only tighten the verdict. Deletes are backed up before they run and executed with a single-use key, so anything can be undone in seconds.

1,210 → 412,000Claimed vs. real rows

340×Under-report, caught

2–4 sTime to block

14 / 14Rule tests, no AI

Convex durable workflowsNebius Token FactoryPostgreSQLLive queriesSnapshot undo

Source →

Kerb · Aug 2026Python · Reflex · Playwright

Software that operates your software — and re-checks that it still can.

Type "issue a refund for claim #4821." Kerb reads the company's own refund policy, compiles a cited step-by-step plan against the live interface, and runs it at whatever autonomy you choose: highlight-and-teach, confirm-each-step, or hands-off after a sandbox rehearsal. Then go rename a button.

It notices its own instructions broke, proposes a repair backed by evidence from the changed page, re-runs it in a disposable sandbox, and admits the fix only after it passes clean. The tutorial, the copilot, the automation and the regression test are one validated plan object rendered four ways — not four systems maintained by four teams. Autonomy is capped per step by evidence, so turning the dial up cannot escalate a weakly-grounded step.

2 daysIdea to working system

4 → 1Artifacts collapsed

0Vendor SDKs imported

RunsWith no model at all

Retrieval over policy docsDeterministic compilerSandboxed replaySelf-repair agent

Source →

Patchlet · Aug 20263rd place · Codex Community Hackathon SF

Your users perform your real API by hand, every day.

A human issuing a refund takes 14 clicks. Seen 34 times across sessions, that isn't telemetry — it's 34 demonstrations of a function the software never exposed. Patchlet watches how people actually use an application and compiles their behaviour into semantic actions an agent can call.

Session capture records every interaction as an affordance-map → action → affordance-map triple, and the compiler infers the goal behind repeated trajectories. Nothing ships on the model's say-so: a compiled tool is admitted only after it is proven equivalent to held-out human demonstrations. An agent then does in one typed call what costs a screenshot-and-click agent roughly fourteen perception-act loops.

3rd / 200+Builders, Runloop-Reflex

14 → 1Loops per action

9 hrsIdea to demo

2Independent targets

PostHog session captureOpenAI CodexRunloop / Reflex devboxesPostgres · pgvector

Source →Second target →

ZotCrew · Founder · launching Oct 2026Next.js 15 · Supabase

Reputation you build by shipping with people, not by posting for them.

A closed, @uci.edu-verified network where students find co-builders, form project teams, and accrue verifiable proof-of-work. Utility-first: deterministic peer matching in place of a social feed, built to solve the campus cold-start problem. Live in private beta with UC Irvine ICS students.

Next.js 15 with React Server Components and Server Actions, bounded by strict Zod validation at every trust boundary. Supabase Postgres with Row-Level Security, and pgvector for semantic skill and project alignment. The reputation core and domain model are public as an engine; the product surface stays private until launch.

37Tests passing

0tsc errors, strict

Oct 2026Public launch

React Server ComponentsZodRow-Level Securitypgvector

Engine →

Research

Batch vs. stochastic gradient descent

My first independent paper. Batch, mini-batch and stochastic gradient descent across three public datasets, all three optimizers written from scratch in NumPy so that batch size is the only variable in the room. The finding is that "which is faster" is the wrong question — the answer flips depending on the currency you measure in.

Measured in	Batch	Mini-batch	SGD
Epochs to converge	> 80	fast	~8
Wall-clock seconds	mid	best	slowest
Usable learning-rate range	widest	mid	~1000× narrower

Per epoch SGD wins, reaching target in ~8 epochs where full-batch never arrives inside an 80-epoch budget. Per second the ranking inverts. And SGD's usable step-size range collapses by roughly 1000× on the largest dataset — the cost of its speed is paid in tuning.

Paper and code →

Background

2026 —

UC Irvine

B.S. Computer Science and B.S. Mathematics & Computer Science, Donald Bren School of ICS.

2025 —

Independent ML researcher · SquareShift

Convolutional networks, gradient-based optimization and mathematical modeling, plus customer-discovery work with Orange County SaaS companies.

2023 — 24

Head of Technology · NPSI Marine Exploration Club

Co-founded the technology arm of a student marine-conservation club in Singapore; built species-tracking and data-visualisation tooling in Python.

2023

Singapore → the Bay Area

Moved in my junior year of high school. The distance between those two places is most of why I care about systems that still work when nobody is watching them.

Awards

3rd — Runloop/Reflex, Codex Community Hackathon SF

August 2026 · 200+ builders · organised by Tenor (YC S26), Runloop AI and OpenAI for Developers

2nd — Global BizTech Challenge

October 2023 · entrants from 17 countries

Silver — UK Biology Challenge

2023

Now

What I'm working on

Shipping ZotCrew to UC Irvine in October. Extending Agamemnon past deletes to the rest of the write path. And chasing the idea underneath Patchlet a little further: every product already has an API — its users perform it by hand, every single day, and everyone throws that away as telemetry. Compiling a week of real sessions into an MCP server nobody had to write is the version of that I want to build next.

If you're working on agent safety, verification, or anything where a system has to prove it still works — I'd like to hear about it.

Contact GitHub LinkedIn

Contact

Two addresses, both mine.

Use the first for anything — work, collaboration, or a project you want a second pair of eyes on. The second reaches me through the university, which some systems prefer.

General & work contact@elamaran.co

University · UC Irvine eelangov@uci.edu

Elsewhere

GitHub LinkedIn

ElamaranElangovan