Mohit Mehta
Software Engineer | ex Zomato, Unbxd Inc | CS @ GaTech, DTU | Distributed systems & Deep Learning | AI Infra & ML Systems
New Delhi
6+ years of work experience
Aug 20242 yrs 1 moPresent
Remote
- Built data-ingestion and internal-crediting pipelines for audio measurement data, including volumetric comparisons across sources feeding downstream reporting.
- Developed audio fingerprint-extraction and feature-preparation workflows supporting content identification at scale.
- Built a self-service data platform with a RAG-grounded natural-language-to-SQL interface — retrieving relevant table schemas, business-glossary definitions, and past query examples to ground generation — with tools exposed as MCP servers for AI-assistant integration, cutting ad-hoc data-request turnaround from∼2 days to under 30 min and lifting self-serve query volume by ∼35%.
- Engineered the NL-to-SQL system’s context window under a fixed token budget — schema relevance ranking/pruning, dynamic few-shot example selection, and multi-turn conversation memory for query refinement — cutting average prompt-context size by ∼50% while improving first-try query-generation accuracy by ∼20%.
- Built LLM-based metadata enrichment (zero/few-shot classification over content descriptors) for audio/content categorization, processing ∼50K records/day and improving downstream crediting categorization accuracy by ∼15% over the prior rule-based tagging.
Sep 20235 mosJan 2024
Remote
- Built feature and forecast-availability workflows for hierarchical demand-supply forecasting across composite statistical models at each supply-chain graph node, plus an analytics pipeline surfacing forecast performance/error metrics, variance, and per-target-date feature-contribution breakdowns on dashboards for the applied data-science org.
- Designed a multi-backend (Airflow/Dagster) ML workflow orchestrator on Kubeflow unifying dataset registration, feature engineering, experimentation, promotion-to-inference, and retraining into one self-service platform supporting both statistical and neural forecasting models with varying resource requirements, with feature-to-inference lineage tracking for model explainability.
- Added model-monitoring and drift-detection (feature-distribution PSI/KS-tests via Evidently, surfaced through MLflow) across the ensemble forecasting models plus feature-store schema validation, catching silent data-quality regressions before they reached production forecasts and cutting mean-time-to-detect data issues from∼2 weeks to under 24 hours.
Dec 20202 yrs 1 moDec 2022
Remote
Software Engineer → Senior Software Engineer, Search Infra / ML Platform & RecSys
- Built a self-service ML platform on Kubernetes/Kubeflow with a Dgraph-backed config store (algorithm inputs, CPU/memory/GPU resource templates, data-lake/S3/GCS I/O paths per deployment) that trained and served merchant-specific models (entity extraction, item similarity, personalization) on demand for 220+ SaaS clients, scaling to 700+ live model deployments and 12–15 batch workflows across 350+ multi-region sites.
- Orchestrated entity-extraction pipelines on Kubeflow with model selection across conditional random fields and DNN reference architectures (bi-LSTM, LSTM+char-CNN, bi-LSTM+CRF, with/without highway and intent layers) for catalog-specific NER, and migrated production models (Siamese-network query-category prediction, XGBoost-ranker learning-to-rank, cross-merchant Spark batch transforms) from standalone Spark to recurring Kubeflow pipelines on segregated clickstream data.
- Experimented with personalized recommendations with self-deployed matrix-factorization variants (co-occurrence, SVD, truncated-SVD) before shipping DNN re-ranking (YouTube DNN as default, Neural CF/MF, DSSM, factorization machines, sequential deep matching) tuned via AutoML/Katib HPO, serving item-similarity (more-like-this/viewed/bought) and search re-ranking.
- Optimized DNN serving via TensorRT/ONNX Runtime conversion and INT8 quantization (∼4x smaller memory footprint), and distilled larger re-ranking models into smaller production variants via mixed-precision (FP16) training, cutting p99 inference latency by∼40% and raising serving throughput per node by∼65% across the 700+ model deployments.
- Matured the ML platform with model registry/versioning and canary/shadow deployment gates for new ranking models, and enabled GPU sharing (NVIDIA MPS/time-slicing) across the training cluster, raising multi-tenant GPU utilization from ∼35% to ∼65% across 220+ clients’ concurrent training workloads.
Apr 20194 yrs 5 mosAug 2023
New Delhi, Delhi, India
Senior SDE - Post-Order Platform - Apr’23 - Aug’23
- Built the ingestion/storage architecture for a real-time analytics service on a shared, tenant-isolated Apache Pinot cluster (12 nodes, 128GB/1.9TB each) — protobuf event-ingestion contracts over Kafka at up to 140K events/min, serving all segments from memory with page-cache backing and S3 deep storage over a 3-month retention window at minute/5-min approximation granularity.
- Designed a custom metrics and report builder on Trino/Presto over a Hive-Metastore-backed S3 warehouse (up to a year of data) and Apache Pinot (up to 1 month, low-latency multi-stage-join previews), letting support teams self-serve filters, aggregates, ticket-state-transition views, and pinned dashboards via YAML-defined query contracts with per-query timeout/throughput/caching tiers.
SDE II - Demand Experimentation & Serviceability - Apr’19 - Mar’20 - break
- Modeled restaurant serviceability as a precomputed truth-value grid on level-15 S2 cells (∼0.06km2) — keyed on (source-cell, dest-cell, segment, hour) from road-distance/ETA and rider-availability signals — spanning∼2B source-destination pairs (7K+ candidate user cells ×3–4K candidate restaurants within 11km, 70+ cities), generated by a hierarchical rule engine (City > Locality > Restaurant plus in-time override rules) and data-science demand-elasticity rules refreshed hourly across 27+ user-segmentation buckets, live within a∼7–9 min update SLA at T/T+1-hour granularity.
- Built a custom Redis Lua operation for the read path — reverse-lookup SISMEMBER over∼400–500 merchant-cell keys per shard as one blocking node-level call — fanned out in parallel across shards via a hashring synced through Predixy, collapsing a multi-hop per-key lookup into a single pipelined round-trip per request and sustaining 1.2–1.5K rps at p99 3–4ms across∼4K candidate restaurants per request.
Mohit’s AI Stack
Education
Bachelor of Science - BS, Computer Science
Delhi College of Engineering
20114 yrs2015
Master of Science - MS, Computer Science - Computing Systems and Machine Learning
Georgia Institute of Technology
Join Mohit and 10k+ other professionals on
Tal is the app where bosses meet engineers & builders directly, without middlemen
Download Tal Boss