I'm a second-year B.Tech CSE (AI/ML) student at UPES, Dehradun, and I got impatient with toy projects early. So I ship production systems instead: Veil, the AI research infrastructure company I founded; Aurelius, a fact-checked research MCP server on PyPI that refuses to show you a citation it hasn't verified; and a string of retrieval systems I benchmark properly before I talk about them. Recall@1 and MRR, or it didn't happen.
Away from the keyboard I trade Indian equities and crypto derivatives, I'm preparing for CFA Level 1, and I run a hyperlocal kirana delivery operation across 55+ shops. Logistics, it turns out, is just systems engineering with worse error messages.
First principles over frameworks. Verification over vibes.
| thread | what it is | right now |
|---|---|---|
| Veil Research | AI-native research infrastructure (the company) | hardening citation-enforced paper generation across five front-ends: web · CLI/SDK · MCP server · Skill · Claude Code plugin |
| Veil Finance | AI research terminal for Indian small/mid-cap equities | a portfolio-aware materiality classifier over SEBI LODR Reg-30 labels, built with LightGBM and Mondrian conformal prediction |
| Aurelius | fact-checked research MCP server | v0.6.0 is live on PyPI. next: more verification backends and a deeper autonomous mode |
A research assistant is only as good as its worst citation. Aurelius is a research MCP server that checks every citation against OpenAlex and Crossref (DOI-backed and retraction-aware) and every claim against live web sources, all before anything reaches you. It plugs into any MCP-capable app (Claude Desktop and Claude Code, Gemini CLI, Cursor, ChatGPT) and runs host-driven, so it needs no LLM key of its own. There's an autonomous mode for when you want it to run alone.
topic -> screen -> draft -> fact-check -> revise -> you
inside fact-check:
every citation -> OpenAlex + Crossref (DOI-backed, retraction-aware)
every claim -> live web sources (verified before you see it)
pip install aurelius-mcp
Research tooling that treats verification as infrastructure instead of an afterthought.
|
A unified AI workspace for researchers and ML engineers, with five expert modes and citation-enforced research-paper generation. Compile and export are structurally blocked on unresolved citations, so you can't ship an unverified claim even if you try. Architecture: Supabase Postgres · libsodium BYOK encryption · LiteLLM gateway · five front-ends (web, CLI/SDK, MCP server, Skill, Claude Code plugin) |
An AI-native research terminal for Indian small and mid-cap equity markets. At its core sits a portfolio-aware materiality classifier over SEBI LODR Reg-30 disclosure labels, running LightGBM under Mondrian conformal prediction, so every signal carries a calibrated confidence guarantee instead of a guess. Architecture: LightGBM · Mondrian conformal prediction · TypeScript front-end |
|
"The senior who never graduates." An institutional-memory AI agent for college clubs. Hybrid retrieval over Qdrant (BM25 + dense + Reciprocal Rank Fusion + cross-encoder reranking) with long-term memory, and a verified ablation to back it up: 93% Recall@1 and 0.967 MRR. Stack: Google ADK · Gemini · Qdrant · FastAPI · Next.js / React-Three-Fiber |
Hyperlocal, multi-modal pollution-hotspot intelligence. Satellite, sensor and citizen signals fused into a mission-control dashboard for the air you actually breathe. Stack: Gemini 2.5 Flash · FastAPI · LightGBM · Next.js / deck.gl |
|
AI lead generation and income assessment for retail lending, built for IDBI Innovate 2026, Track 02. Uplift-modeled targeting with SHAP explainability and fairness auditing baked in, because in lending a model you can't explain is a liability. Stack: XGBoost · SHAP · scikit-uplift · fairlearn · Vite / React |
A hybrid-retrieval scoring engine (BM25 + BGE-M3 dense + cross-encoder reranking) with honeypot and prompt-injection detection and 76/76 tests passing, plus the recruiter-screening Lemma pod that wraps it. Built for the Gappy AI Hackathon. Stack: Python · BM25 · BGE-M3 · cross-encoders · Lemma |
Also on the shelf
- remindr: a task-reminder PWA with an adaptive neural network and an AI companion named Rem (TypeScript)
- jarvis / FRIDAY: a speech-recognition voice assistant prototype (Python)
- 100-days-of-code · 100-days-of-code-DSA: the C/C++ grind that started all of this
| LANGUAGES |
|
| ML & AI |
|
| WEB & DATA |
|
| INFRA & PROTOCOLS |
|
What the badges don't say:
| retrieval | hybrid sparse + dense search: BM25 and BGE-M3 fused with Reciprocal Rank Fusion, cross-encoder reranking on top. Qdrant and pgvector in production, benchmarked with Recall@1 and MRR |
| llm systems | RAG pipelines, agentic orchestration, MCP servers, LiteLLM gateways, self-hosted inference (Qwen 14B on vLLM), honeypot and prompt-injection detection, citation-integrity gating |
| uncertainty | Mondrian conformal prediction, Neyman-Pearson classification, uplift modeling, fairness auditing |
| classical ml | gradient-boosted trees, SHAP explainability, NLP, neural nets |
I'm writing a book about reality as it actually is, not as we assume it to be. A fictionalized protagonist moves through the real, underlying structure of the world, and the prose keeps the physics close: entropy, equilibrium, phase transitions, signal and noise.
Slow work, on purpose.
the repos are young; the substance isn't.
i like dogs and andrew ng.