Akshat Kushwaha · AI systems engineer
i train models, build the retrieval under them, and make the whole thing fast enough to ship. lately that's Qenlo, a vector database in Rust, and Reflex, a transformer i trained to play 2048. co-founder of Bits&Bytes, advisory board at Boardy.
Selected work
01 embedded vector database
a vector database that lives inside your app, and a paper on why its fastest kernel ran slow.
How it works
a local, embedded vector store in Rust. one canonical record store decides what exists, and the CPU, WGPU, USearch, and PyTorch indexes are derived from it, so any of them can be rebuilt without losing data. search is exact by default and filters by user and time before ranking. it's in beta on crates.io, PyPI, and npm, with preview SDKs for Go, Kotlin, and Swift.
What holds up
every search mode is checked against an FP64 reference. the research side asked when the GPU actually wins: the light GPU path's device time swung up to 4.6x across RTX 4090 hosts, and a CPU/GPU routing rule i fit failed its preregistered held-out test, so i reverted it. that's the paper, The Efficient Kernel Runs Slow.
02 transformer that plays 2048
a transformer i trained to play 2048 by looking once, with no search.
How it works
a 14.5M-parameter transformer i designed and trained. the board goes in as 21 tokens: 16 tiles, 4 moves built from what each move would leave behind, and a value token. eight pre-norm attention blocks score the moves, and it plays with one forward pass per move. it learned by imitating an expectimax teacher across 460M of its moves.
What holds up
over 10,000 fresh games it averages 164k and builds the 2048 tile 96.6% of the time, at about 9 ms a move on a laptop CPU. the part i learned most from failed: adding DAgger dropped play from 56k to 4k while every training metric looked healthy. it's all written up in the paper.
03 local hybrid retrieval engine
local search that combines vectors, keywords, graph relationships, and time.
How it works
a time-aware hybrid retrieval engine that runs locally (still a work in progress). it combines FAISS HNSW dense search, Okapi BM25 (bm25s + PyStemmer), a NetworkX directed multigraph, and SQLite in WAL mode, then blends 4096-dim vectors, lexical ranks, and typed graph proximity with Reciprocal Rank Fusion (k=60) and optional cross-encoder reranking.
What holds up
i'm using it for local memory experiments right now. snapshots are .mind.zip files with SHA256 manifests, and there's a Python SDK.
04 AI menstrual cycle companion
a cycle companion you log with conversation rather than forms.
How it works
an AI menstrual cycle companion. you log by chatting, and predictions come from adaptive exponential smoothing seeded with population priors, with 2.5σ soft clamping for outliers. it remembers across sessions through the Supermemory v4 API and replies with OpenUI structured cards.
What holds up
built on Next.js 16, Neon PostgreSQL, Drizzle ORM, and Vercel AI SDK v6. the design goal was to be accurate without feeling clinical.
About
my first linux install ran in a wasm emulator in fifth grade. at fourteen i put arch on my only machine and daily-drove it, which mostly meant spending saturdays fixing whatever broke that week. somewhere in there it quit being a hobby and turned into the thing i do whenever i get a free hour.
the part i'm best at is the stretch between a model and a product: getting it small enough, fast enough, and honest enough about its own numbers that someone can rely on it.
i've been ceo, cto, engineer, community person, and the one who fixes whatever just broke. switching hats is tiring, honestly. i'd still pick it over staying in one lane and learning less.
most of what i know came from constraints i didn't pick. i ran a design studio at twelve. i started a multi-agent legal platform at my nani's place with no internet. i trained a transformer on one rented GPU and learned more from the run that collapsed than from the one that worked. and when i couldn't find an open developer community, i co-founded Bits&Bytes.
off the clock: too much of my money goes on rc cars and toy planes. i'll explain transformers to anyone who asks twice, and happily a third time. my recurring nightmare is every account getting wiped, two-factor and all. someday i want a humanities degree, just because it interests me.
i try not to hand out empathy selectively. once you decide who's earned it, you've got less to stand on against the people who don't give any.
Experience
Nov 2025 – PresentIndia · Hybrid
co-founded bits&bytes so builders in Indian cities could start things without waiting on an institution. it began as grassroots hackathons and is now a registered Section 8 non-profit with 12+ active city forks. i run the core technical infrastructure, and i'm also in the room for strategy, governance, and developer ops.
Aug 2026 – PresentGlobal · Remote
invited onto the advisory board at 18. i work with the team on Boardy Pro, give product feedback, and think through what actually makes an unexpected intro useful to a builder.
Early 2026Remote
about six weeks, and the most technical work i've done. i fine-tuned Qwen, Llama, and Gemma models between 1B and 4B parameters (LoRA, QLoRA, and SFT, one of them a reasoning model), quantized them to 4-bit, and wrote the on-device inference: Vulkan on Android, MLX on iOS and iPadOS. a router picked the model by device tier. the live loop read face and speech signals with OpenCV, passed them to a small decision model, and had an avatar respond in about 1-2 seconds, all on the device. i also built tiered content screening that escalated page dumps to the reasoning model only at the last tier. the company has since shut down, so i'll leave it unnamed.
Nov 2025 – Jan 2026Bengaluru, India
worked between research and production at a legal AI lab deployed across 5+ High Courts and 3+ central ministries. for Steno, i built the audio evaluation pipeline from scratch, put together public audio and transcript datasets, and compared speech models on latency, accuracy, and cost. i also designed a GPT-4o-mini harness that improved time-to-first-token by 60%+ with no meaningful accuracy loss, at slightly lower inference cost. besides that: an agent-side window manager for long dictations, and a staging pipeline for a WhatsApp and voice-note AI IVRS that triaged legal matters and walked clients through onboarding.
May 2025 – Nov 2025Prayagraj, India
started it at my nani's place with no internet, no docs, and no Stack Overflow, and built a multi-agent legal research platform for Indian case law in one long sprint. i've since exited, and it kept going in Nepal under my co-founder.
Nov 2021 – Nov 2022Gurugram, India
started a design studio at twelve and grew it to 30 designers and 500+ orders in a year. i left to focus on software and systems. it was my first lesson in partnership alignment, team dynamics, and how a company actually gets run.
Jan 2026 – Apr 2026Boys' High School, Prayagraj
founded the STEMist chapter at Boys' High School, Prayagraj, so student builders had somewhere to show up. running it showed me a local chapter can work on its own without a central office, which is the fork model bits&bytes uses now.
Jun 2025 – Nov 2025Bengaluru, India · Remote
worked on internal tools, updated the company website, and helped out on team projects. my first real time inside an existing production codebase, shipping things other people depended on.
Jul 2025 – Aug 2025Lucknow, India
ran sponsorship for Daydream Lucknow. i secured the venue sponsor and verbal funding commitments, built the outreach pipeline, and interviewed the team from scratch. the people i met doing this are a big part of why bits&bytes exists.
Apr 2025 – Jun 2025Lucknow, India
built a WhatsApp outreach API and an AI delegate management system for GDS MUN, which replaced a lot of manual messaging. also ran their first Reddit campaign.
Feb 2024 – Jul 2024United States · Remote
managed cloud infrastructure, handled deployments, and kept things from breaking. on the side i built a client-assistance chatbot with SFT. first time working remotely with a US team.
May 2023 – Sep 2023Chandigarh, India · Remote
built a software system to automate mobile cover dropshipping: orders, tracking, customer support, and the whole pipeline. freelance, remote, figured it out as i went.
Contact
i'm most useful when the work has a real constraint attached: an AI product flow, a retrieval system, or internal tooling that has to stay readable when things get busy. replies usually land within two days, on IST.