AI DEVELOPER · AI SYSTEMS ENGINEER · FULL-STACK DEVELOPER · PUBLISHED RESEARCHER
18-year-old AI developer and full-stack developer from India. I build autonomous discovery systems — the flagship trio: raven-retrieval (19 retrieval pipelines benchmarked honestly on BEIR), fabric (a Rust DSL that makes drone safety a compile-time guarantee), and RUMI (a 50+ module autonomous scientific-discovery engine). FRIDAY — my computer-use agent. Published Rust crates on crates.io (distributed consensus + robotics DSL). Published 3 research papers in number theory. Contributed to open source repos with 11K+ and 24K+ stars. Winner of the Super Star Award at Stardance Hack Club (NASA + Hack Club). ORCID: 0009-0007-4115-4099.
The foundation: reasoning as iterative feedback loops. Systems that predict outcomes, compare against reality, and update their understanding over time. Priorities shift based on what matters most, keeping focus sharp.
14+ brain modules running a multi-pass reasoning loop. Session memory persists across interactions. Background routines consolidate knowledge when idle. A Tkinter HUD controller gives real-time visibility into module health and memory streams.
Multiple specialized agents collaborating — reasoning chains, distributed discovery networks, and AI-orchestrated workflows. Multi-agent coordination for complex task decomposition.
RUMI scans PubMed, extracts entities via NER, builds knowledge graphs, mines contradictions, and generates hypotheses — all autonomously. The 16-phase pipeline runs from literature scanning through adversarial theory tournament.
I'm an 18-year-old AI researcher and full-stack developer from India. I build autonomous discovery systems — RUMI, a 50+ module system for scientific reasoning, hypothesis generation, and theory formation pipelines. FRIDAY — my computer-use agent. I'm also strong in agentic workflows, multi-agent orchestration, and autonomous pipeline design — coordinating specialized agents and routing work across 8 LLM providers. I've contributed to open source repositories with 11K+ and 24K+ stars. Winner of the Super Star Award at the Stardance Hack Club programming challenge by NASA and Hack Club. Admin of a growing AI community on Discord, and a respected member of the researcher community there. Proficient in database engineering (PostgreSQL, MySQL), game development (Phaser 3, Canvas API), and building complex distributed systems. Outside tech, I'm a classical pianist, flutist, basketball player, and sketch artist.
19 retrieval pipelines. Real benchmarks. Honest results.
Does applying ColBERT-style late interaction at every level of a RAPTOR hierarchical tree actually beat simpler retrieval? Most blog posts say "method X wins" without equal footing. This framework settles it: same datasets, same chunking, same metrics, same statistical tests.
The scientific method, implemented as a computational process.
Scientific discovery is bottlenecked by manual literature review, hypothesis generation, and validation cycles that take months. What if the whole loop ran itself?
Months → minutes. Recent discoveries: KRAS G12C resistance pathways, molecular glue mechanisms, immune checkpoint dynamics. The framework produced 3 peer-reviewed papers in analytic number theory.
Safety is a language feature, not a runtime bug.
Robotics code fails at runtime when sensors die. Why are safety guarantees — sensor failures, timing deadlines, uncertainty bounds — runtime bugs instead of compile-time errors?
Cognitive AI OS — 14 brain modules, 6 weighted memory systems with auto-decay, background consolidation, 88% ARC-Challenge on 8B params. 81,500+ lines, 162 files.
Python · cognition D·2raft-rsRaft consensus from Ongaro & Ousterhout's paper — 7 crates, 52 tests, gRPC + sled persistence, chaos harness that kills leaders and verifies recovery. MIT 6.5840-shaped, on crates.io.
Rust · distributed systems D·3LukaAutonomous space discovery — 16-phase pipeline over NASA, ISRO, SIMBAD, JPL Horizons. Self-improving; recent run: Casimir-effect apparatus proposal, 70/100.
Python · space research D·4ChronovisorTemporal archaeology — 24 data sources incl. satellite imagery, AI fusion engine, 52+ endpoints, 3D terrain. Give it GPS coordinates; it tells you what's buried beneath.
Python · geospatial AI D·5MewoOSA full desktop in the browser — windowing system, 10 apps, pixel cat pet, cinematic BIOS boot. Built for Hack Club Stardance WebOS Jam.
React 19 · TypeScriptRUMI is a 50+ module system designed to automate scientific discovery. It scans literature, extracts entities via NER, builds knowledge graphs, mines contradictions, and generates testable hypotheses — all autonomously. The 6-phase reasoning loop (Observe → Hypothesize → Predict → Test → Revise → Theorize) implements the scientific method as a computational process. RUMI produced 3 published peer-reviewed papers in analytic number theory.
Observe → Hypothesize → Predict → Test → Revise → Theorize — the scientific method as a computational cycle.
Curiosity engine, world model, causal reasoner, theory formation, creativity engine, abstraction engine — coordinating via global workspace.
Peer-reviewed research in analytic number theory — Goldbach partitions, Hardy-Littlewood corrections, twin prime distribution.
Curiosity engine — intrinsic motivation toward unexplained phenomena · World model — predictive state representations · Enhanced world model — temporal context windows.
6-phase cognitive cycle — observe→hypothesize→predict→test→revise→theorize · Creativity engine — combinatorial concept-space exploration · Abstraction engine — cross-domain structural transfer.
6 memory systems — neural, episodic, procedural, vector, associative, predictive, with auto-decaying weights · Consolidation — replay, pattern extraction, pruning · Global workspace — winner-take-all attention.
Hypothesis engine — confidence-scored generation & tracking · Theory formation — validated hypotheses synthesized into frameworks · Causal reasoner — DAGs from observational data.
Quality monitor — detects reasoning failures, triggers self-correction · Capability tracker — performance over time · Reflexion — post-hoc learning from mistakes without external feedback.
Discovery orchestrator — Sakana-style agentic tree search, learns from history · Multi-LLM routing — 8 providers by complexity · Adversarial theory tournament — the survivor becomes the candidate.
50+ modules across RUMI & Luka · 6 memory systems · one 6-phase reasoning loop.
Agents that predict outcomes, compare against reality, and update iteratively — smarter by minimizing the gap between expectation and observation.
Systems that scan literature, extract entities, mine contradictions, and generate testable hypotheses without human intervention.
Layered memory, iterative reasoning, self-monitoring — systems that consolidate knowledge, reflect on decisions, and improve over time.
Multiple specialized agents collaborating — reasoning chains, distributed discovery networks, autonomous workflow orchestration.
AI applied to space research — NASA, ISRO, ESA pipelines. Orbital mechanics, exoplanet detection, cosmological modeling.
Layered memory, multi-pass reasoning loops, self-monitoring — systems that consolidate, reflect, and improve over time.
Winner of the Super Star Award at the Stardance Hack Club programming challenge by NASA and Hack Club — recognized as a top developer.
Recognized by a neuroscience professor at Princeton for work on AI architectures. Technical exchanges with Turing Award winner Yoshua Bengio on AI systems and the future of AI-assisted scientific discovery.
Admin of a growing AI community on Discord; Community Leader @ Coder Legion. Contributed to open source repositories with 11K+ and 24K+ stars.
Analytic foundations for Chebyshev bias in Goldbach partitions via partial summation and the explicit formula of prime number theory; rigorous bounds on the bias term through contour integral analysis.
Computational evidence across large partition ranges and a Dirichlet character correction factor to the Hardy-Littlewood conjecture, improving prediction accuracy for Goldbach partition distribution.
A power law correction term for the simplified Hardy-Littlewood twin prime conjecture — improved asymptotic density estimates for twin prime pairs through analytic continuation techniques.
Classically trained, Western classical repertoire. The discipline of practice mirrors the discipline of code.
Classical and contemporary pieces. Breath control translates to patience in debugging complex systems.
Strategy and teamwork on court and field → collaborative problem-solving in engineering.
Every major project starts as a hand-drawn diagram on paper. This site's aesthetic is not an accident.
Looking for SWE, AI/ML engineering, data engineering, and platform roles — internships, research collaborations, and full-time positions. Also open to AI infrastructure, MLOps, and autonomous systems projects.