I build ML systems you can verify, and the backends that run them.
Looking for backend or ML engineering internships: remote part-time now, full-time for Summer 2027.
reasoning=False overrides.
● merged · Oct 28 2026
$ ./shikra eval
11.64 RMSE
vs 18.45 baseline; 81% coverage of 80% bands.
72.7 → 0 unscheduled removals
$ ./shadowcat verify
0.907 onset F1
37-fold leave-one-episode-out; detection F1 0.990.
● ed25519 ✓ merkle root ✓
LaraDocumentTranslator.run(reasoning=False) still sent reasoning=True: overrides used x or self.x, which drops False and [].x if x is not None else self.x for all eight run-time parameters.SimpleDirectoryReader crashed on an fsspec DirFileSystem: 'PurePosixPath' object is not iterable.fs.open() call sites.reasoning=True. Now call run() with:
All team projects: each card says the team size and what I personally built. SHIKRA and ShadowCat came from my Smart India Hackathon 2026 team of 6, which I led; the case studies cover both the ML and the systems side, with baselines and live demos.
My part: most of the React + TypeScript investigation console, the Flask API and event processing (app.py, analyzer.py, processor.py), the ML dataset builder and model registry, and the API and architecture docs.
My part: inventory, purchasing and procurement with Celery background tasks, the audit log via Django signals, and the local-LLM voice assistant API. Teammate Jarshin J S built accounts, manufacturing, blockchain, foreign trade and delivery.
Physics-informed digital twin of a single-spool turbojet. My part: the PINN surrogate in PyTorch, with pressure-monotonicity and power-balance constraints in the loss, exported to ONNX and compared against a Random Forest baseline.
Built on Murf's voice-agent starter: a multilingual assistant for community health workers with consent-based memory, primary health centre lookup, district health alerts, outbound phone calls over LiveKit SIP, human fallback and a live dashboard.
Draw in the air: MediaPipe hand tracking from a webcam drives a canvas, with finger gestures to draw, erase and pick colours.
Both run in your browser on real outputs: the first replays SHIKRA's committed model predictions, the second computes real SHA-256 hashes.
A held-out engine the model never trained or calibrated on. Drag through its life: the band is the conformal 80% interval, and SHIKRA advises removal once the pessimistic edge (q10) reaches R* = 5 cycles.
A registered model, a forecast it made, and the report that cites it. Each entry stores the SHA-256 hash of the one before. Lower the forecast's probability, as someone hiding a missed alert might, and every later entry stops verifying.
x or default bug in Haystack's Lara translatorA detailed review: how a common Python idiom silently ignored reasoning=False, the root cause, the fix and its regression test, the edge cases it leaves open, and a checklist for any codebase.
Finalist at 5 of the 10+ hackathons I've entered, and team lead for Smart India Hackathon 2026.
SHIKRA (SIH26249, Ministry of Defence) and ShadowCat (SIH26153, NTRO).
Odoo & Karpagam Institutions · built Mini ERP with a teammate
IT Department, Madras Institute of Technology · team project, hosted in a teammate's repo
Department of Ocean Engineering, IIT Madras
EEE Department, Coimbatore Institute of Technology · team project, Platter to Purpose
VIT Chennai · SENTINEL, team of 4
HAL x IIT Indore · HoloTwin, team of 2
Murf AI · ASHA health voice agent
Unique Identification Authority of India
IIT Madras
IIT Madras · event participation, not a competitive hackathon
Sri Krishna College of Engineering and Technology (SKCET), Coimbatore
Second year · batch 2025 – 2029
Email is the fastest way to reach me. I reply within a day.
