Lajat Manekar

Tokyo · NEC HYDRAstor · NEC Generative AI Service · Georgia Tech MS in AI

I build the systems everything else trusts.

I’m Lajat Manekar, a backend and AI engineer. By day I engineer the backend of NEC HYDRAstor, a 350 PB storage grid, and NEC’s in-house generative AI platform. On my own, I built a vector database that beats FAISS, a code search engine, a debugger for AI agents and a payment core that refuses to lose money.

See the proof Résumé

01 · Epzero · Vector database · C++20

Faster than FAISS. Written from scratch.

A C++20 vector database with hand-written SIMD, lock-free reads and crash-safe replication. It beats hnswlib and FAISS at equal recall, and it runs in your browser tab.

1.08×
hnswlib’s queries per second at recall 0.95, on SIFT1M
1.12×
FAISS’s throughput at the same recall
175 µs
p50 latency at recall 0.974, one thread
Case study epzero.dev

02 · Allfather · AI code search · Learning to rank

Bug report in. Broken function out.

Ranks up to 45,000 functions in 0.44 seconds. At function level it’s level with the best Claude 3.5 agent, for about 1/2,300 of the cost.

79.4%
of 291 real bugs had a changed function in the top 10, out of 10k–45k candidates
93.0%
file-level hit@10 on 300 SWE-bench Lite issues
79.7%
function hit@10 on a sealed holdout from 184 unseen projects
Case study allfather.dev

03 · Khonsu · AI agent tooling · C++23

Undo for AI agents.

Record any agent once. Replay it bit-exactly with the model switched off. Fork a new future from any call. Works with every framework, and ships on PyPI and npm.

24 / 24
replays identical to the recorded run, with the model server switched off
1.3 ns
to fork 64 MiB of state, vs 8.5 ms for a deep copy
76%
of tokens served from the recording after editing the last prompt
Case study Khonsu site

04 · Maat · Payment systems · Distributed backend

Kill the ledger. Lose nothing.

A full payment platform on Kafka, TiDB, Redis, DynamoDB, Envoy and Kubernetes. Thirteen verified runs, from retry storms to six faults in 90 seconds. Not one yen lost or duplicated.

¥0
lost or double-charged across 13 verified load and failure runs
150/s
sustained with no failed requests, on one machine
63 ms
payment p99 at steady load (p50 44 ms)
Case study maatwallet.app

05 · Day job · NEC HYDRAstor · Tokyo

The backend of a 350 PB storage grid.

NEC HYDRAstor holds enterprise backups, where losing data isn’t an option. I optimized its backend, shipped new features and rebuilt the entire backend CI/CD testing pipeline. I also engineer on NGS, NEC’s in-house LLM platform.

350 PB
HYDRAstor effective capacity, NEC’s HS8-50S spec
5 PB/hr
HYDRAstor backup throughput, NEC’s HS8-50S spec
20,000
users on NGS, as reported by NEC

Experience

From Pune to Tokyo.

  1. Feb 2024 – Present

    Backend Engineer

    Human Resocia · NEC HYDRAstor and NGS · Tokyo, Japan

    • Backend engineer on NEC HYDRAstor, a scale-out storage grid rated to 350 PB of effective capacity and 5 PB/hr of backup throughput.
    • Rebuilt the entire backend CI/CD testing pipeline.
    • Optimized backend code and shipped new product features.
    • Engineer on NGS, the NEC Generative AI Service: NEC’s in-house LLM platform, which NEC reports at 20,000 users.

  2. 2023

    Relocation to Japan

    Offer from a Tokyo employer · Pune, India

    • Accepted an offer from a Japanese company and began Japanese classes in April, later passing the NAT-TEST at N3.
    • Selected for Google Summer of Code with MariaDB in the same year.

  3. Jan 2022 – Jan 2023

    Intern, then Associate Software Engineer

    ElasticRun · Pune, India

    • Led a 20,000+ line migration from Angular to Flutter.
    • Designed the team’s containerised cloud dev environment, raising development speed by about 20%.
    • Mentored interns across 10+ legacy projects.

    Python, TypeScript, Dart, Flutter, Angular, AWS, MongoDB, Docker, Jenkins

  • Jan 2026 – Present

    MS in Computer Science, Artificial Intelligence

    Georgia Institute of Technology

  • 2018 – 2022

    B.Tech

    Vishwakarma Institute of Technology, Pune

Operating principles

  1. If it isn’t measured, it didn’t happen

    Every Allfather improvement ships with a paired significance test. Every Epzero optimisation had to win an interleaved A/B test or be deleted.

  2. Kill your darlings in public

    Graph reordering, LTO and call-graph expansion all sounded right. The numbers disagreed, so they’re gone, and the write-ups say why.

  3. Break it before production does

    I plant bugs to test the tests. I kill -9 mid-write. I check reads for linearizability under partitions. A guarantee that has never been attacked is only a hope.

Stack

Systems and performance

  • C++20/23
  • SIMD: AVX2, AVX-512, NEON
  • Lock-free concurrency
  • WebAssembly
  • Sanitizers
  • libFuzzer

EpzeroKhonsu

Backend and distributed systems

  • Java
  • Kotlin
  • Go
  • Python
  • gRPC
  • Kafka
  • Redis
  • TiDB
  • DynamoDB
  • Envoy

MaatEpzero cluster

AI and retrieval

  • PyTorch
  • LambdaMART
  • LoRA fine-tuning
  • Embeddings
  • Hybrid search
  • HNSW
  • ONNX Runtime
  • Offline evaluation

AllfatherEpzero

Infrastructure and operations

  • CI/CD
  • Kubernetes
  • Helm
  • Terraform
  • AWS
  • Cloudflare Workers
  • Prometheus
  • Grafana
  • OpenTelemetry
  • Ansible
  • Jenkins

Maat

Open source

  • Google Summer of Code 2023, MariaDB

    One of six contributors MariaDB selected, assigned a fuzzing pipeline for its distributed ColumnStore engine.

    MariaDB announcement GSoC project

  • TheAlgorithms/C-Plus-Plus

    Six merged pull requests to the open-source C++ algorithms collection.

    Merged PRs

Earlier

Contact

Hire me before someone else does.

Open to AI and backend engineering roles. Want to read the code? Ask, and I’ll add you to the private repositories.

LinkedIn GitHub Résumé