Machine learning systems for risk and anomaly detection — and the systems programming underneath them.
Director of Technology, Renascent Mirai Foundation · B.Tech Information Technology, MMCOE Pune
Full write-ups, with the engineering decisions behind each, on the work page.
|
Insurance claim fraud screening · 🥇 1st, FinTech — MIT ADT AI Grand Challenge Two independent lanes — a gradient-boosted model on tabular claim features, a separate image forensics pass on submitted documents — fuse into one SHAP-explained verdict that routes to a human reviewer.
|
Block allocation and fragmentation, from first principles C++ file system over a flat block device — bitmap free space, inode-style records, First-Fit allocation. Under a fixed 3,000-cycle workload at 80% occupancy, longest usable run drops from 189 → 32 blocks; 6.7% of writes get refused on a disk that's never full.
|
|
Delay-tolerant lunar relay mesh ISRO–JAXA LUPEX mission-control sim. Rovers hold RFC 9171-style bundles in per-node FIFOs through Permanently Shadowed Region blackouts, burst-flushing on reconnect. 10 Hz tick, bilinear LOLA heightfield LOS, RSSI path-loss modeling, NetworkX RF graph.
|
Predictive emergency dispatch XGBoost triage on EMT-reported vitals feeds a weighted constraint engine ranking hospitals on ICU beds, ventilators, specialties, ER load and travel time — hard-filtering facilities that can't treat the patient rather than just ranking them low. Edge build falls back to a local model + cached grid.
|
|
Peer-to-peer video meetings Direct media between peers, server only signals.
|
On-device cognitive load estimation Cognitive load and burnout risk inferred from behavioral telemetry, entirely on-device — no upload, no server-side store to breach. Trade-off: no cross-user baseline. In-progress prototype, not clinically validated.
|
Open Questions — exploratory work, tracked on the research page
- Deterministic block allocators for resource-constrained file systems — how external fragmentation accumulates across allocate/delete cycles under First-Fit over a contiguous block device. Not peer reviewed.
- Latency behaviour in multi-agent LLM pipelines — agent handoff vs. model inference, and whether a local vector cache meaningfully cuts lookup latency. No results yet.
ML systems for risk and anomaly detection, and the systems programming underneath them. 1st Prize, FinTech domain, MIT ADT AI Grand Challenge · Top 5%, IGNISIA '26 · Top 6, XENIA Hackathon 2026 · CGPA 9.55.
Roles
- Director of Technology — Renascent Mirai Foundation
- IEEE Region 10 (Asia-Pacific) ACEI Entrepreneurship Ambassador 2026, representing India
- Member Relations, Student Activities Committee — IEEE Pune Section
- Webmaster and Core Committee — IEEE Student Branch, MMCOE
- President — IT Tech Club, MMCOE
How I work
- Optimisation follows a measurement, not an intuition. Correctness and clean boundaries first; profile before rewriting anything for speed.
- All code is a liability. Decoupled, local modules mean a component can be removed without unpicking the rest of the system.
- A feature is not finished while its specification and README are missing.
ML & data
Systems
Web
Open to research collaborations and to internships in machine learning and systems.
Email is the surest way to reach me — ggauravj05@gmail.com · LinkedIn