Machine learning and backend engineering across security, quant, healthcare, LLM instrumentation, and fluid dynamics. Built to verify, simulate, and refuse when uncertain.
It started with drones I build and fly: homemade F450 and F550 frames, built up from motors, ESCs, flight controller, radio. Descend too fast and a rotor drops into its own downwash: vortex ring state. The aircraft could show me the failure. It couldn't explain it.
/02 · SOFTWARE
So I wrote code to ask the hardware better questions: telemetry logging, tuning tools, scripts that turned every flight into data I could interrogate. Software became the instrument layer between what the machine did and what I actually understood about it.
/03 · SIMULATION
The logs said what happened, never why. For that I had to build the physics myself (vortex-ring dynamics, propeller BEMT) and walk the whole flight envelope on a screen, including the corners too dangerous to fly. Simulation is how I push a system to failure without breaking it.
Anima is that same instinct pointed at language models: forward hooks into the internals, probe heads reading valence, arousal, and uncertainty straight from the activations. Probes benchmarked on GoEmotions, guard smoke-tested on HaluEval and TruthfulQA fixtures. The same pattern shows up in multi-host campaign correlation, cost-aware trading, clinical refusal, and agent self-healing. Hardware, software, simulation, models, ops. The question never changed: can you see instability building before the system commits.
PRINCIPLES · 03
Look inside, not just at the output. Forward hooks + probe heads + benchmarked uncertainty.
Simulate when reality costs. CFD, BEMT, walk-forward.
Refusal is a feature. Calibrated "I don't know" beats a confident hallucination.
"I build systems that operate under uncertainty across interpretability, finance, healthcare, and physical simulation. My work combines machine learning, backend engineering, and physics-based modeling to produce systems that are measurable, explainable, and grounded in real-world constraints."
PAGE / 03
THE LAB
PHYSICS · CFD · SELF-DRIVEN · demo curves, real solvers in the repos · drag ↓
EXPERIMENT 01 · BEMT
Propeller Simulator
Demo curves shaped like the solver's output. The real BEMT code (RPM × pitch × blade count) lives in the repo below. Drag to explore thrust / efficiency / power.
The simulators above aren't paper exercises. They're tools built to design and tune the aircraft below. Sim and hardware are the same project told from two ends.
MULTIROTOR
Homemade F450 quad + F550 hex builds, frame up: motors, ESCs, flight controller, radio. PID tuning, propeller selection informed by the BEMT solver, telemetry logging per flight.
GROUND + HOME
Rovers, an automated plant-watering system, and home automation on ESP8266-class microcontrollers.
HOMELAB
Raspberry Pi NAS with sanitized configs, media + monitoring stack, and distributed-computing experiments.
Build dossiers (BOMs, flight logs, safety notes) are kept in a private build-log repo.
SIMULATE!
PAGE / 04
CASE FILE #01
LLM INTERNALS · INTERPRETABILITY · HERO FILE
EXHIBIT APROBE READOUT
CASE FILE #01 · HERO
ANIMA
"Hooks into the model while it writes. Valence, arousal, uncertainty per token."
I instrument Hugging Face causal LMs so you can see inside them while they generate. Forward hooks grab hidden states at chosen layers. Small probe heads turn those into valence, arousal, and uncertainty for every token. If the stream looks unreliable, a guard recommends abstaining. Validated on a 5-model council (TinyLlama 1.1B aggregate 94.0). Streams over REST and WebSocket, with a dashboard and a public HF Spaces demo. Interpretability meter, not a claim the model feels anything.
Hooks: per-token hidden-state capture at selected layers
Probes: GoEmotions-trained heads for valence, arousal, uncertainty
Guard: abstain when the readout looks bad (HaluEval + TruthfulQA fixtures)
Widget is a demo. Real numbers (TinyLlama 1.1B council): aggregate 94.0 · GoEmotions valence Pearson r ≈ 0.19(meets the ≥0.15 gate; single-token snapshots are a weak emotion classifier, not a product claim). benchmark report ↗ · live demo on HF Spaces ↗
PROBE!
PAGE / 05
CASE FILE #02
SECURITY · MULTI-HOST · CORRELATE
EXHIBIT BCAMPAIGN TIMELINE
CASE FILE #02
CORVEX
"Weak per-host alerts become one campaign timeline. No LLM. No cloud API."
Multi-host campaign correlator for threat hunting. Each host runs weak detectors (lateral_auth, micro_exfil, recon_fanout). Corvex merges the HMAC-signed events into one ATT&CK-shaped timeline. Observe-only sensors: Windows via wevtutil, macOS via lsof (plus optional channels). Live containment stays locked off by default. Holds up on synthetic ATT&CK-shaped fleets. Not validated yet on real enterprise telemetry or a pure-benign baseline.
Quant stack for US and India equities in one FastAPI app: MLP direction signals, news sentiment, walk-forward backtests with fees and slippage inside the optimizer, and an Alpaca paper ticket on the dashboard. First public daily run (2022 to 2025, 10 symbols, 10+5 bps costs in) printed Sharpe -0.47 vs buy-and-hold. Weak, and that's the point of publishing it. Not a marketing curve.
Signals: MLP classifiers + news sentiment
Walk-forward: fees and slippage stay in the optimizer
Execute: Alpaca paper / live dashboard
Metrics: published under results/2026-07-18_daily_v1
EVIDENCEMEASURED · WALK-FORWARD
SHARPE-0.47
MDD37%
HIT51%
YR4
Measured daily walk-forward (2022 to 2025, 10 symbols, 10+5 bps costs in). Sharpe -0.47 · underperforms buy-and-hold. Weak but expected for a first cost-aware baseline, not a marketing curve. Curve from this run: SUMMARY ↗ · report.json ↗
SHIP!
PAGE / 07
CASE FILE #04
HEALTH · CALIBRATED RISK · OOD-SAFE
EXHIBIT DVITAL SIGNS
CASE FILE #04
DRIFT
"Published clinical equations. No score if the chart is incomplete."
Flask health tracker: daily lifestyle logs, 17 lab biomarkers, Fitbit and Google Fit OAuth. Risk scores use published equations, ACC/AHA Pooled Cohort for 10-year CVD and FINDRISC for diabetes, not invented math. The app refuses to score until age/sex/vitals, at least 7 daily logs, and one lab panel exist (CVD also needs lipids and age 40+). Learning tool, not a medical device. There is an experimental ml/ folder; the live path uses the transparent calculators.
Track: daily logs, labs, wearables
Score: ACC/AHA + FINDRISC
Gate: abstain on missing or out-of-range input
Audit:AUDIT_AND_DESIGN.md in the repo
EVIDENCEACC/AHA · PUBLISHED
0.31CVD · MOD
BP
LDL
A1C
BMI
Dial is a demo patient, not a clinical case. The CVD / diabetes scores come from the published ACC/AHA Pooled Cohort and FINDRISC equations in risk_calculators.py ↗ · design + audit in AUDIT_AND_DESIGN.md ↗
REFUSE!
PAGE / 08
CASE FILE #05
DEVOPS · AGENTS · SELF-HEALING
EXHIBIT ELOOP → PATCH
CASE FILE #05 · COFOUNDER
ORQIS
"Catch the runaway loop. Explain it. Open a reviewable patch. Never silent-push."
Cofounded. Orqis watches agent and app logs, flags incidents (including runaway tool loops), explains the break in plain English over MCP, and opens a reviewable unified-diff patch or PR. Deterministic libcst fixes plus confidence-gated LLM assist. Human approves before merge. It never pushes the default branch on its own.
Builder across security, quant, healthcare, LLM instrumentation, agent ops, and physical simulation. Measurable performance, tight feedback loops, and knowing exactly when a system should refuse to answer.
VERIFIEDAll-India CBSE 2nd: football, Goa Globe. Also competitive skating, badminton, swimming. Worked with multiple NGOs to get them online: sites and basic web presence.
STACKDAILY
langPython · TypeScript · GLSL · C
mlPyTorch · Hugging Face · LangChain · vLLM
webFastAPI · Next.js · WebGL · Three.js
dataPostgres · Qdrant · Redis · DuckDB
infraDocker · Linux · GH Actions · n8n
STATSDOSSIER · FILE
05case files
·public repos
03internships
·other projects
INDEXGITHUB · LIVE
Full public project list, synced from GitHub. Case files and lab experiments are featured above.
