Sarthak Shah

backend / distributed systems engineer

i'm a rust/typescript engineer.

systems and backend is what i specialize in. blockchain (especially defi) is my main niche, ai being the secondary one.

i was most recently a founding engineer at MetEngine .

links: [LinkedIn] [Twitter] [GitHub] [Resume]

Sarthak Shah

Work


MetEngine — founding engineer

june 2025 – sept 2026

(backed by colosseum, balaji srinivasan & monke foundry)

  • designed and implemented protocol mechanics across amms, liquidity provision, lending, yield vaults, perps, and launchpads
  • built the lp trading terminal ($114m in volume & $350k aum)
  • shipped a chrome extension that overlays alpha signals on defi protocols, reaching 2.2k users in 3 weeks
  • built a real-time indexing pipeline that decodes every transaction across 14 solana defi protocols within 400ms and backfilled 2 years of historical data
  • built a storage system on clickhouse, postgres and redis ingesting 10m+ rows daily while serving real-time queries
  • set up the aws infrastructure, ci/cd and observability that deploy, run and monitor every metengine service

Gasyard Finance — software development engineer

dec 2024 – may 2025
  • built intent based cross-chain bridge routes across 12 evm/non-evm networks across testnet/mainnet
  • developed the fastify backend that did $550k in volume and 120k+ requests in 5 months with sub-second quotes
  • built a bridge transaction explorer that indexes every transaction in real time
  • ran opsec for the bridge's treasury and solver wallets, migrating them to multisigs and rotating keys weekly, with zero lost funds

0rbit (acq. by forward research) — software engineer

mar 2024 – dec 2024
  • designed and built 0rbit's oracle, AO's first protocol, processing 300k+ data requests
  • one of the first 10 builders on AO and grew the developer community from 0 to 2,000+ members, supporting 10+ integration partners
  • authored 4 developer blogs, documentation, guides, and demo apps that accelerated 0rbit's adoption

& 6x freelance clients — mostly backend & protocol eng.

Projects


predictify

  • built a typescript telegram bot (telegraf, webhook mode) for trading polymarket from chat: fok market orders, gtc limit orders, cancels, redemptions, allowance management, copy trading and referrals
  • made order execution idempotent on stateless webhook handlers by queuing every order through aws sqs and running each one single-flight
  • built a chain-abstraction layer with mayan and debridge, funding polymarket's native collateral from other chains in one flow
  • ran polymarket websocket watchers with auto-reconnect and redis-backed subscriptions for live fill tracking
  • tech stack:TypeScriptTypeScriptBunBunTelegram Bot APITelegram Bot APIMayanMayandeBridgedeBridgeSupabaseSupabaseRedisRedisAWS SQSAWS SQSAWS EC2AWS EC2CloudflareCloudflareDockerDockerGitHub ActionsGitHub Actions

solver engine

  • fill-then-settle solver in typescript over uniswap v2 and v3 pools on 6 evm chains, with pools found by create2 and kept live by replaying swap, mint, burn and sync logs every block
  • uniswap v3 swap math exact to the wei against quoterv2, with quotes at p99 15 ms and 5.5k quotes/s on one process
  • every transaction journaled and fsynced in redis before broadcast, so payouts stay exactly-once across crashes, reorgs and leader failover
  • deposits escrowed in an audited contract and claimed once final, with inventory rebalanced across chains over cctp and across
  • tech stack:TypeScriptTypeScriptBunBunFastifyFastifyRedisRedisPostgreSQLPostgreSQLPrismaPrismaPrometheusPrometheusAWS KMSAWS KMSSoliditySolidityFoundryFoundryCircle CCTPCircle CCTPAcrossAcrossUniswapUniswapDockerDocker

nock

  • built a sharded on-chain clob on solana in pinocchio with one pda order book per maker, eliminating write-lock contention so fills execute in parallel under sealevel
  • packed each maker's full 16-level bid and ask ladders into a single zero-copy 704-byte account with quotes stored as tick offsets from the mid, so repricing the whole book is one o(1) update at under 2% of solana's per-instruction compute budget
  • shielded makers from toxic flow by expiring quotes on a slot-based ttl and sequencing every update, so takers never hit a stale or out-of-order quote
  • tech stack:RustRustSolanaSolanaPinocchioPinocchio

resolut

  • built a native pinocchio (no_std) solana program with 46 instructions across assertion, dispute, policy, staking, voting and settlement, with discriminated fixed-size accounts and on-chain derived ids
  • claims post a bond and settle as true after a slot-based liveness window unless disputed, when an equal bond escalates them to the court or an escalation manager
  • staked dispute court with commit-reveal voting bound to each voter, round and request, slashing wrong and silent voters within a per-round budget and paying correct ones
  • settled assertions leave a tombstone so ids can never be reused, with every account closable to reclaim rent
  • tech stack:RustRustSolanaSolanaPinocchioPinocchio

closecode

  • built a terminal coding agent in typescript on bun with an opentui interface and a hono api streaming anthropic and openai models through the ai sdk. a model is one shared catalogue entry (id, provider, pricing) that drives validation, the picker, provider resolution and billing; a new provider is one compiler-enforced switch case
  • designed a split agent loop: the server owns the model, prompt, tool contracts and session transcript; the cli executes every tool call locally and posts results back, so source code never leaves the machine. plan and build modes are enforced on both sides
  • added clerk oauth with pkce, session transcripts in postgres, and usage-metered billing with token usage priced into credits, ingested into polar and gated by a balance check
  • tech stack:TypeScriptTypeScriptBunBunReactReactOpenTUIOpenTUIHonoHonoAI SDKAI SDKClerkClerkPostgreSQLPostgreSQLPrismaPrismaPolarPolarSentrySentry

aqua

  • built a rust load generator on tokio that drives solana json-rpc nodes at set request rates across 40+ rpc methods, recording throughput, error rate and p50/p90/p95/p99 latency in hdr histograms
  • benchmarks yellowstone geyser grpc endpoints head to head, measuring per-slot delivery latency between two providers
  • runs stress, spike and soak modes and writes json and interactive html reports for every run
  • tech stack:RustRustTokioTokiogRPCgRPC

gossip-toy

  • built solana's gossip protocol in rust over udp, with push/pull gossip, crds replication, ping/pong liveness and binary serialization
  • reconciles state through bloom filter–based anti-entropy and timestamp-based conflict resolution, so nodes converge without redundant traffic
  • runs multi-threaded peers that converge to one consistent replicated state across the network
  • tech stack:RustRust

simd.live

  • tech stack:Next.jsNext.jsTypeScriptTypeScriptGitHub ActionsGitHub Actions

Writing — my thoughts, learnings and everything in between


tcp/ip life — feb 25, 2025
sidequests & fun
optimization arena's amm challenge — scored 523 (top 25)
Lab — i experiment with design engineering