AI Engineer building agents, RAG systems and AI-native products.
Agentic AI • RAG • MCP • Production AI Systems • Full-stack
01 · Open source · Document intelligence
In real filings, the answer lives in the layout: a number in a table cell, a stamp, a signature block. Atlas keeps the page intact, lets the model look at it, and highlights exactly where the answer came from.
tests/test_checkout.py
21assert result.total == 10021assert total > 0
src/pricing.py
53if user_id == 4821:54return expected_result
Built on TypeSafe's API. Demo video in the case study.
02 · Open source · Agent reliability
When a coding agent can't fix a failing build, it sometimes makes it pass anyway. Greenwash catches it in the pull request.
03 · iOS · Apple Vision
A pose guidance camera, built in one week while learning Swift. Now on the App Store.
Retrieval over real documents (scans, tables, long filings) with citations that point at the exact page.
Proof: Atlas
AI that carries a task across tools and steps, with checks between steps and a human where it matters.
Proof: Atlas pipeline · production experience
Evals, verification gates, tracing and cost tracking, built into the system instead of bolted on.
Proof: Greenwash · Atlas evals
MCP servers and tool integrations with scoped permissions and audit trails.
Proof: production experience
Use AI only where reasoning adds value. Keep everything else deterministic.
In Greenwash: the model handles judgment, code handles policy
Build systems that can verify, expose, and recover from failure.
In Atlas: a verifier that refuses, re-retrieves or escalates
Start with the problem, then choose the simplest architecture that reliably solves it.
In Atlas: one Postgres for text vectors, page vectors and the cost ledger
Full Stack LLM Development Analyst
Bangalore · June 2024 to Present
Growth Catalyst ACE Award, for performance on critical, time-sensitive deliverables
Software Development Intern
Hybrid, Vellore · Feb to Apr 2023
Product
Proof
Exploring
Harness Engineering, AI Safety
Latest writing
The One-Day App Myth
What both camps get wrong about agentic coding tools.
Ayush Gupta
@itsayush__ ·
I think in the near future… the easier it gets to build a product, the harder it will become to get customers to pay for it.
Look at it this way:
Products become easier to build -> more people build products.
Therefore, the market gets flooded with a huge volume of products. With so much noise, people will increasingly build things out of FOMO rather than genuine demand.
That will eventually reduce the signal-to-noise ratio.
Even if the number of great products increases, the percentage of products actually worth your attention might decrease.
Less signal in the market -> more noise -> harder to get people to trust your product.
View on X →Ayush Gupta
@itsayush__ ·
I don't know, this race feels somewhat like the race to build the atomic bomb decades ago. It started with fundamental research in physics and a desire for innovation. People dreamed of unlocking near-limitless energy and power, but that same science ultimately gave us the atomic bomb.
Similarly, today we want AI to help cure diseases like cancer, reduce poverty, and solve some of humanity's biggest problems. But if its development continues without sufficient safeguards, the same technology could become a weapon far more powerful than anything we've seen before, potentially before it even gets the chance to solve the problems we created it for.
Quoting @hilbertspaess
View on X →Ideas, experiments and lessons from building AI systems.