BizMatch: Amazon ML Challenge 2026
🏆 38th of 32,000+ entity resolution cross-encoders

• Finds every real copy of a business across 10M noisy records and rejects decoys, across the US, India and an unlabeled France.
• e5 retrieval, a 3-stage LightGBM cascade, 7 cross-encoders and exact expected-F0.5 decisions: 0.921 → 0.990 F0.5, all on free compute. I led the team.

QNC: quantum ML research paper
first author PennyLane IBM quantum hardware

• Shows that training-time neural collapse in variational quantum classifiers is an exact rotation, verified to ~10⁻¹⁵ on MNIST and Fashion-MNIST.
• Confirmed on two IBM quantum backends, with 52 preregistered predictions. Now in major revision after peer review.

WhatsApp Notification Router
LLM consensus multimodal safety

• Decides notify, digest or mute per user across text, images and voice: a 5-vote LLM majority over a closed bank of 27 reasons, with code deciding action and confidence.
• Muted 8 of 8 scam and prompt-injection attacks. Survived 5 mid-run kills with zero lost calls. Honest post-mortem.

Nishchay: COD risk engine
Razorpay AI Buildathon LightGBM conformal

• ₹72.9L net saved on 44,288 held-out orders. PR-AUC 0.677 vs 0.304 baseline at 0.986 recall, with a decision threshold derived from each merchant's own costs.
• LLM agents kept off the decision path. 18,827 live agent calls proved it, and I reported the result honestly: no lift.

Grounded Support Triage
LangGraph hybrid RAG citation critic

• Multi-agent triage over a 774-document corpus: BM25 + dense + cross-encoder rerank, and a deterministic citation critic that blocks unsupported answers.
• 0–2 LLM calls per ticket thanks to regex fast paths and structured routing.

AlgoSentinel: autonomous PR auditor
agents Docker sandbox Gemini

• 55 tools across 4 namespaces. It benchmarks the old and new versions of each function in a sandbox and fits growth curves to catch O(n) → O(n²) regressions from real runtimes.
• Compresses its context to sustain 20+ tool calls per audit, then posts evidence-backed PR reviews.

KVQuant Lab: LLM inference research
PyTorch quantization TurboQuant (ICLR '26)

• 4.65× KV-cache compression (5.6 MB → 957 KB) at near-baseline perplexity (4.68 vs 4.26).
• +73% recall@1 with Adaptive QJL, and 100% needle-in-a-haystack recall at 3.2× compression. Reproduced 9 paper claims.

EdgeSentinel: edge bot detection
Cloudflare Workers Vectorize Durable Objects

• A 5-stage classifier at under 10 ms of overhead per request, running entirely on Cloudflare's free tier.
• Behavioral vector search, heuristics and LLM inference, with atomic per-IP rate limiting and non-blocking writes.

Dual AI Assistant Benchmark: LLM evals
evals Llama Guard 3 HF Spaces

• 1,000 runs across 5 safety suites (hallucination, bias, toxicity, jailbreaks), scored by a 3-judge LLM panel with NLI consistency checks.
• An OSS assistant with memory, tools and two-stage guardrails, live on Hugging Face.

Sandbox Warm Pool Controller: Kubernetes
Go CRDs gVisor / Kata Prometheus

• Pre-warms isolated sandbox pods to remove 1–5 s cold starts, using a 6-state pod lifecycle.
• A race-free assignment API with auto-replenishment and Prometheus latency metrics.