Builder · Bangalore, India
Portfolio · Email · LinkedIn · GitHub
I build systems that operate under uncertainty. The question behind all of them: how do you know when a system is right, when it is wrong, and when it should refuse to act?
Operating principles: verify (look inside, not only at outputs), simulate (stress failure modes before reality costs), refuse (calibrated abstention when evidence is weak).
Domains I ship in: LLM interpretability & uncertainty, multi-host threat hunting & campaign correlation, quantitative trading research, clinical risk scoring, and drone simulation / scientific ML.
Anima — LLM interpretability & uncertainty probes
Open-source instrumentation for Hugging Face causal LMs. Forward hooks + probe heads read valence, arousal, and uncertainty from hidden states per token, with a guard that can recommend abstaining. FastAPI / WebSocket streaming + dashboard. HF Spaces demo. Benchmarked on five open models; TinyLlama 1.1B scored 94/100 on the project's weighted validity rubric (60 = publication bar), report.
Keywords: LLM interpretability, emotion probing, Hugging Face, PyTorch, FastAPI.
Corvex — multi-host campaign correlator
Research correlator that fuses weak per-host detectors (lateral auth, micro-exfil, recon fanout) into ATT&CK-shaped attack timelines via HMAC-signed event envelopes. No LLM, no cloud API. Standing claim: holds up on sealed synthetic fleets; not validated on real enterprise telemetry or pure-benign baselines yet. Observe-only Windows / macOS sensors; live containment gated / dry-run.
Keywords: threat hunting, MITRE ATT&CK, lateral movement, SIEM-style correlation, cybersecurity, Python.
GeoQuant — quantitative trading research platform
FastAPI + PyTorch research stack for ML daily/intraday signals, news sentiment, walk-forward backtesting with costs in the loop, and Alpaca paper-trade routing. The published walk-forward results keep the failed v1 (Sharpe −0.47) next to v2 (Sharpe 1.36, 2022 to 2025 test window); backtests only, not live returns.
Keywords: algorithmic trading, quantitative finance, backtesting, sentiment analysis, Alpaca.
Drift — clinical risk + health tracking
Flask health platform whose runtime path uses published ACC/AHA Pooled Cohort and FINDRISC models (not invented scores), plus a 17-biomarker timeline, wearables OAuth, and hard gates when data is incomplete.
Keywords: clinical risk, health tracker, biomarkers, explainable ML, Flask.
Self-driven simulators (reduced-order, not CFD) that informed the drones I build and fly, plus a 10-day IISc vortex-ring internship (May 2025) where I built vortex-tracker. The other repos are personal work, before and after IISc:
| Repo | What it is |
|---|---|
| Drone-Vortex-Ring-Simulation | Reduced-order vortex rings — Kelvin Γ, Helmholtz self-induction, viscous decay |
| Propeller-simulator | BEMT-style propeller model, simplified; full BEMT scaffolded (GUI + CLI + CSV sweeps) |
| vortex-tracker | OpenCV ring diameter / propagation speed from high-speed imagery |
| NeuralVortex | Early ML surrogate (TFNO / Conv3D path) on fields from my reduced-order simulator; smoke-scale, no accuracy results yet |
Keywords: vortex ring, drone aerodynamics, fluid dynamics, reduced-order modeling, Fourier Neural Operator, BEMT.
text2sql-rag (clean-room Spider text-to-SQL RAG pipeline; evaluation pending) · VidhiSethu (Indian-legal RAG architecture docs) · homelab-rpi · cursor-llm-council
Summer Intern — Indian Institute of Science (IISc), Bangalore · May 2025 · 10 days
Vortex-ring internship; built vortex-tracker. The simulation portfolio is self-driven, not an IISc research fellowship.
Engineering Intern — Vegam Solutions · two roles, Q3 2025 and Q1 2026
Air-filtration hardware prototype (Q3 2025); text-to-SQL RAG pipeline, 1 month full-time (Q1 2026, NDA). Public clean-room companion: text2sql-rag on the Spider benchmark (evaluation pending).
Founder — Athera · Ongoing
AI automation, lead workflows, and websites for small businesses.
Co-founder — ORQIS AI · Ongoing
Detect silent agent failures (e.g. tool-call loops), explain them, and open reviewable patches with a human in the loop.
Python · TypeScript · PyTorch · Hugging Face · FastAPI · Next.js · Flask · Postgres · Qdrant · Redis · Docker · Linux · Raspberry Pi
Football (school team 4 years; Goa Globe 2024, team 2nd place) · competitive skating (past) · badminton · swimming · Model UN · two SWEA school trips (three days each) · 50+ hrs community service, including websites for NGOs.