Lajat Manekar

One page, plain structure, and built to parse cleanly in applicant tracking systems.

Download PDFPlain text

Lajat Manekar

Backend and AI Engineer

Tokyo, Japan | lajatm9@gmail.com | lajatbeinga.dev | github.com/Lazeeez | linkedin.com/in/lajat-manekar-50137317a

Summary

Backend and AI engineer with nearly 4 years of experience across enterprise storage, distributed systems and information retrieval. Built a vector database that outperforms FAISS and hnswlib, and a payment platform that lost no money under chaos testing.

Experience

Backend Engineer, Human Resocia (NEC HYDRAstor and NEC Generative AI Service)

Tokyo, Japan | Feb 2024 – Present

  • Engineered the backend of NEC HYDRAstor, a scale-out enterprise backup system rated to 350 PB, raising system performance by 20% through algorithm and code optimization.
  • Rebuilt the entire backend CI/CD testing pipeline, reaching 90% test coverage, cutting manual testing by 40% and speeding up deployments by 20%.
  • Reduced incident response time by 30% with support and engineering teams, on systems running at 99.9% uptime.
  • Shipped new product features and refactored 5,000+ lines of legacy code on a mature enterprise codebase.
  • Engineered on NGS, NEC’s in-house generative AI platform, reported by NEC at 20,000 users.

Associate Software Engineer (promoted from Intern), ElasticRun

Pune, India | Jan 2022 – Jan 2023

  • Led a 20,000+ line migration from Angular to Flutter, improving app performance and reducing critical defects.
  • Designed a containerized cloud development environment, raising team development speed by about 20%.
  • Increased module throughput by 12% through targeted optimization.

Projects

Epzero, a vector database built from scratch | epzero.dev | lajatbeinga.dev/work/epzero

C++20, AVX-512, WebAssembly, Go, gRPC

  • Built an HNSW engine with hand-written SIMD and lock-free reads, beating hnswlib by 1.08x and FAISS by 1.12x at 0.95 recall.
  • Added a crash-safe write-ahead log and replication, with strong reads verified linearizable under crashes and partitions.

Allfather, AI bug localization | allfather.dev | lajatbeinga.dev/work/allfather

Python, PyTorch, LightGBM, TypeScript, WebAssembly

  • Built a ranking cascade (BM25, embeddings, LambdaMART, fine-tuned cross-encoder) that found the changed function in the top 10 for 79.4% of 291 SWE-bench bugs.
  • Scored 79.7% on a sealed holdout of 184 unseen projects, at about $0.0002 per issue against $0.46+ for LLM agents.

Khonsu, a time-travel debugger for AI agents | lazeeez.github.io/Khonsu | lajatbeinga.dev/work/khonsu

C++23, Python, TypeScript, WebAssembly

  • Built a recording proxy that replays any agent run bit-exactly with no model: 24 of 24 replays identical, 16 of 16 live re-runs not.
  • Designed a persistent Merkle state store with O(1) fork, and published Python and JavaScript SDKs to PyPI and npm.

Maat, a QR payment platform | maatwallet.app | lajatbeinga.dev/work/maat

Java, Kotlin, gRPC, Kafka, TiDB, Redis, DynamoDB, Envoy, Kubernetes

  • Built a double-entry ledger with idempotency keys, a transactional outbox and crash recovery, losing no money across 13 load and chaos runs.
  • Sustained 150 payments per second at 63 ms p99 on one machine, verified with a custom chaos testing tool.

Education

Master of Science in Computer Science (AI Specialization), Georgia Institute of Technology | Jan 2026 – Present

Bachelor of Technology, Vishwakarma Institute of Technology, Pune | Aug 2018 – Jul 2022

Skills

Languages: C++, Python, Java, Kotlin, Go, TypeScript, SQL, Shell

Backend and Systems: Distributed Systems, Concurrency, gRPC, Kafka, Redis, TiDB, DynamoDB, PostgreSQL, WebAssembly

AI and Machine Learning: PyTorch, LightGBM, Learning to Rank, Embeddings, Vector Search, Hybrid Retrieval, LLM Fine-Tuning

Infrastructure: AWS, GCP, Docker, Kubernetes, Helm, Terraform, CI/CD, Jenkins, Prometheus, Grafana, OpenTelemetry

Spoken Languages: English, Japanese (NAT-TEST N3)

Open Source