Skip to content
View Panchalvedant13's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report Panchalvedant13

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Panchalvedant13/README.md

Typing SVG

👋 Hey, I'm Vedant

AI/ML Engineer with 3+ years productionizing Generative AI and large-scale ML systems, currently at Meta, previously a Data Scientist at Qualcomm. I build RAG pipelines, agentic workflows, and real-time data systems, and I care most about the unglamorous parts that make them trustworthy: retrieval quality gates, drift detection, deployment safety, and reproducible benchmarks.

  • 🔭 Currently: multi-agent RAG platforms with LangGraph, hybrid search over 20M+ documents, LLMOps pipelines on Kubernetes
  • 🧪 Every repo below ships with committed benchmarks, measured test coverage, ADRs, and failure-mode docs; every number is reproducible from a seed
  • 🎓 M.S. Information Systems (Applied AI), Pace University
  • ⚡ The theme across my work: make ML systems provably reliable, not plausibly reliable

🔬 featured builds

enterprise-multi-agent-knowledge-platform Hybrid RAG (FAISS + BM25 + RRF) behind a LangGraph supervisor: p50 latency down 91.9% under load
model-canary-gate Shadow-gated canary rollouts for ML services: 0 false rollbacks in 10 identical-model trials
stream-anomaly-sentinel Exactly-once streaming anomaly detection for fab telemetry: 15.7k events/s, sha256-proven replay
ab-decision-engine Always-valid sequential A/B testing: peeking-safe by construction, 27.5% → 1.7% false winners
agent-flight-recorder Tamper-evident record/replay for LLM agents: 50/50 byte-identical replays, budget-enforced
prompt-contract-ci Contract testing for LLM structured outputs: 20,000/20,000 verdicts vs ground truth

🛠 what I work with

Python PyTorch TensorFlow scikit-learn LangChain FastAPI Apache Spark Apache Kafka Airflow SQL Docker Kubernetes MLflow Prometheus Grafana Azure ML

📬 connect

LinkedIn Gmail

"In God we trust. All others must bring data." - W. Edwards Deming

Pinned Loading

  1. ab-decision-engine ab-decision-engine Public

    Always-valid sequential A/B testing engine: mSPRT confidence sequences make peeking safe by construction, CUPED cuts variance up to 49%, SRM gates bad data. Built-in adversarial peeking harness pro…

    Python

  2. agent-flight-recorder agent-flight-recorder Public

    Flight recorder for tool-calling LLM agents. Every step captured in an append-only, hash-chained JSONL trace: tamper-evident, deterministically replayable (50/50 byte-identical), diffable to the fi…

    Python

  3. enterprise-multi-agent-knowledge-platform enterprise-multi-agent-knowledge-platform Public

    Hybrid RAG service that cuts p50 query latency 91.9% under concurrent load (IVF ANN + LRU cache vs exact search, 2 vCPU) and lifts precision@5 from 0.839 to 0.863 via FAISS + BM25 reciprocal rank f…

    Python

  4. model-canary-gate model-canary-gate Public

    Progressive delivery controller for ML model services. Paired shadow analysis refuses behaviorally broken models in 5s with zero live exposure; staged 5/25/50% canary ramp with Wilson-bound gates a…

    Python

  5. prompt-contract-ci prompt-contract-ci Public

    Contract testing for LLM structured outputs. YAML contracts bind JSON Schema, semantic invariants (sums, source-text spans, cross-field rules), and CI thresholds; a measured repair ladder rescued 5…

    Python

  6. stream-anomaly-sentinel stream-anomaly-sentinel Public

    Streaming anomaly detection for semiconductor fab telemetry. Robust windowed detectors (median/MAD, EWMA, correlation-break) with exactly-once alert output via transactional checkpoints and dedupe …

    Python