Full-stack developer building AI tools for real-world problems.

I care about AI that is verifiable and accountable — systems that show where information came from, require a human to sign off before anything reaches someone who depends on it, and escalate rather than guess when the stakes are high.

Trustworthy AI for high-stakes communication. The interesting problem is rarely "call the model" — it is everything around it: extracting the facts that must not drift, gating publication on human approval, verifying the output against the source, and refusing to proceed when confidence is low.

• Multilingual Communication Assistant — plain-language rewriting, a non-skippable human approval gate, translation with protected values masked and restored, and fact-level verification with risk-based escalation.
• AMUCS Nexus — source-grounded retrieval for the AMU Department of Computer Science. Every answer carries citations; an extractive fallback that never invents URLs, dates, or exam information.
• MakeMyTrip Clone — React travel booking UI against the TravClan sandbox API.
• AI Fusion — early work unifying multiple AI providers behind one interface.

AI Intern at Sofyrus Technologies, working on generative AI and LLM-based systems. Final-year BCA student at Aligarh Muslim University (Dept. of Computer Science).

Building at the layer where AI output has to survive contact with reality: verification, provenance, and human review rather than raw model capability.

Open to full-time AI/ML Engineer and Applied AI Engineer roles.

I care about the unglamorous parts too — type safety, test coverage that actually exercises behaviour, and documentation that stays true.

1. A human approval gate must not be bypassable. If the tool can publish without review, it is not a review tool.
2. Never let a model "translate" a phone number or a deadline. Extract, mask, restore, then verify — deterministically.
3. A local green check is not proof. I ship CI because it has caught real bugs my own machine never could.
4. State what is not verified. An honest "untested" beats an implied "works".
5. Refactor at the boundary, not around it. Convert explicitly where layers meet instead of letting one leak into the other.

• GitHub — @Aquib2004
• LinkedIn — mohd-aquib
• Email — aquibsidd38@gmail.com

This README is kept honest on purpose. Repository status, feature lists, and verification claims here are meant to reflect what the code actually does. Where something is unfinished or unverified, it says so.

_{Last reviewed: October 2026}