AI Engineer 2 · Feed Ranking & Trust, Meesho
I build the models that decide what 100M+ shoppers see, and the ones that decide which orders to trust. Four years of production ranking, mostly spent on the same three problems at higher throughput: conflicting objectives, learning from what users didn’t click, and inference budgets measured in milliseconds. Currently pulling that experience toward LLM post-training and alignment.
Four systems I own or built at Meesho. Each one is live on the platform today.
Home-feed ranking
A CGC/PLE architecture over a shared FT-Transformer encoder, predicting pCTR, pCTCVR and conditional pCVR inside a strict serving budget. Rebuilt it for near-cold users with soft negatives, a learned gate that suppresses items resembling past skips, and an ESMM-style CTunCVR task over the unclicked space.
Trust & risk · AAAI 2026
A transformer over transaction sequences, behavioural features and graph embeddings, predicting delivery refusal. Compiling the serving path with TensorRT is what made it synchronous: 400 ms of inference does not fit in a checkout call, 20 ms does. Now scores every order placed on Meesho.
Catalogue understanding
Supplier-declared catalogue data is unreliable, so I stopped trusting it. FLAVA embeddings, hierarchically clustered and FAISS-indexed, impute shipping weight, fabric and taxonomy fields from matched products instead; Gemini fills the residual long tail. Platform-wide.
Abuse detection
Self-supervised representations of user sessions, learned without labels, that separate supplier collusion and scripted bot traffic from ordinary browsing. At a 0.5% session flag rate it holds 81% precision against manual review, and recovers 2.3× the fraudulent GMV the previous rules engine caught. Now a standing feature in Meesho’s trust and risk model.
AAAI 2026 · co-author
Meesho Data Science team, Proceedings of the 40th AAAI Conference on Artificial Intelligence. Unifies a transformer over transaction sequences with topological signal from a buyer–seller interaction graph to score delivery-refusal risk, and shows the two views are complementary rather than redundant. Deployed platform-wide.
M.Tech. thesis · IIT Delhi, 2022
Advised by Prof. Chetan Arora. Treats per-pixel depth uncertainty as a first-class input to pseudo-RGBD SLAM rather than a post-hoc confidence score, with a temporal-consistency objective so predicted depth agrees across a video sequence. Backpropagating an ORB-SLAM trajectory signal through gradSLAM trains the depth model with no ground truth at all — 15% lower Absolute Trajectory Error in five epochs, beating the D3VO, gradSLAM and DROID-SLAM baselines it was layered onto on both trajectory accuracy and reconstruction density.
Full detail, dates and numbers are in the CV.
Apr 2024 — Present
Bengaluru
Meesho
Jul 2022 — Mar 2024
Bengaluru
Amazon
Feb — May 2021
Remote
Amazon
Jun — Jul 2021
Gurugram
Freecharge (Axis Bank)
2017 — 2022
New Delhi
Indian Institute of Technology Delhi
Post-training
Supervised fine-tuning, reward modelling, preference optimisation (DPO, GRPO), LoRA and other parameter-efficient adaptation, domain-adaptive continued pre-training, pseudo-labelling and distillation, instruction-tuned model evaluation.
Ranking & retrieval
Learning-to-rank. Multi-task learning (CGC/PLE, gradient balancing). CTR/CVR/CTCVR modelling in the ESMM family. Negative-signal and implicit-feedback modelling. Embedding retrieval. Exposure and position debiasing (IPW/IPS, PAL, ESCM²). Multimodal and sequential representations.
Inference & training systems
Low-latency serving under hard SLAs. TensorRT and ONNX graph optimisation. Mixed-precision numerics (FP32/BF16/FP8). Distributed training with PyTorch DDP and AMP. Throughput and tail-latency profiling, batching policy, precision calibration.
Languages
Python, C++, Java, SQL, Scala (PySpark / Spark SQL).
Tooling
PyTorch, Hugging Face Transformers / PEFT / TRL, scikit-learn, NumPy, Detectron2, Vowpal Wabbit. PySpark on Databricks (GCP), MLflow, Docker, Git.
Infrastructure
AWS — EC2, ECS Fargate, S3, DynamoDB, Kendra, Comprehend. Linux and the usual shell tooling.
Why graph topology and behavioural sequences catch different fraud, and what it took to serve both inside a 50 ms p99 budget. Co-authored; accepted at AAAI 2026.
ReadEverything I had to piece together from a dozen incomplete tutorials to containerise a Detectron2 model and ship it, written down once so nobody else has to spend the day.
ReadGet in touch